SIX-VERTEX, LOOP AND TILING MODELS: INTEGRABILITY AND COMBINATORICS arXiv:0901.0665v2 [math-ph] 7 Dec 2009 PAUL ZINN-JUSTIN Abstract. This is a review (including some background material) of the author’s work and related activity on certain exactly solvable statistical models in two dimensions, including the six-vertex model, loop models and lozenge tilings. Applications to enumerative combinatorics and to algebraic geometry are described. Contents Introduction 3 1. Free fermionic methods 5 1.1. 5 1.1.1. Definitions Operators and Fock space d action gl(∞) and gl(1) 5 Free fermionic definition 8 1.2.2. Wick theorem and Jacobi–Trudi identity 8 1.2.3. Weyl formula 9 1.2.4. Schur functions and lattice fermions 10 1.2.5. Relation to Semi-Standard Young tableaux 11 1.2.6. Non-Intersecting Lattice Paths and Lindström–Gessel–Viennot formula 12 1.2.7. Relation to Standard Young Tableaux 13 1.2.8. Cauchy formula 13 1.1.2. 1.2. 1.2.1. 1.3. 7 Schur functions 8 Application: Plane Partition enumeration 14 1.3.1. Definition 14 1.3.2. MacMahon formula 15 1.3.3. Totally Symmetric Self-Complementary Plane Partitions 16 1.4. Classical integrability 17 2. The six-vertex model 18 2.1. 19 Definition 2.1.1. Configurations 19 2.1.2. Weights 19 Integrability 20 2.2. 1 2.2.1. Properties of the R-matrix 20 2.2.2. Commuting transfer matrices 21 2.3. Phase diagram 21 2.4. Free fermion point 22 2.4.1. NILP representation 22 2.4.2. Domino tilings 23 2.4.3. Free fermionic five-vertex model 23 Domain Wall Boundary Conditions 24 2.5. 2.5.1. Definition 25 2.5.2. Korepin’s recurrence relations 25 2.5.3. Izergin’s formula 27 2.5.4. Relation to classical integrability and random matrices 27 2.5.5. Thermodynamic limit 29 2.5.6. Application: Alternating Sign Matrices 30 3. Loop models and Razumov–Stroganov conjecture 32 3.1. 32 3.1.1. 3.1.2. 3.2. Definition of loop models Completely Packed Loops Fully Packed Loops: FPL and 32 FPL2 models Equivalence to the six-vertex model and Temperley–Lieb algebra 32 33 3.2.1. From FPL to six-vertex 33 3.2.2. From CPL to six-vertex 33 3.2.3. Link Patterns 34 3.2.4. Periodic Boundary Conditions and twist 35 3.2.5. Temperley–Lieb and Hecke algebras 36 3.3. Some boundary observables for loop models 38 3.3.1. Loop model on the cylinder 38 3.3.2. Markov process on link patterns 38 3.3.3. Properties of the steady state: some empirical observations 39 3.3.4. The general conjecture 40 4. The quantum Knizhnik–Zamolodchikov equation 41 4.1. 41 Basics 4.1.1. The qKZ system 41 4.1.2. Normalization of the R-matrix 43 4.1.3. Relation to affine Hecke algebra 43 4.2. Construction of the solution 44 4.2.1. qKZ as a triangular system 44 4.2.2. Consistency and Jucys–Murphy elements 46 2 4.2.3. Wheel condition 47 4.2.4. Recurrence relation and specializations 48 4.2.5. Wheel condition continued 49 Connection to the loop model 50 4.3. 4.3.1. Proof of the sum rule 50 4.3.2. Case of few little arches 51 4.4. Integral formulae 54 4.4.1. Integral formulae in the spin basis 54 4.4.2. The partial change of basis 54 4.4.3. Sum rule and largest component 55 4.4.4. Refined enumeration 57 5. Integrability and geometry 57 5.1. 57 Multidegrees 5.1.1. Definition by induction 57 5.1.2. Integral formula 58 5.2. Matrix Schubert varieties 58 5.2.1. Geometric description 58 5.2.2. Pipedreams 59 5.2.3. The nil-Hecke algebra 60 5.2.4. The Bott–Samelson construction 62 5.2.5. Factorial Schur functions 62 5.3. Orbital varieties 63 5.3.1. Geometric description 63 5.3.2. The Temperley–Lieb algebra revisited 65 5.3.3. The Hotta construction 66 5.3.4. Recurrence relations and wheel condition 66 5.4. Brauer loop scheme 67 5.4.1. Geometric description 67 5.4.2. The Brauer algebra 69 5.4.3. Geometric action of the Brauer algebra 70 5.4.4. The degenerate limit 71 References 73 Introduction Exactly solvable (integrable) two-dimensional lattice statistical models have played an important role in theoretical physics: starting with Onsager’s solution of the Ising model, they have provided 3 non-trivial examples of critical phenomena in two dimensions and given examples of lattice realizations of various known conformal field theories and of their perturbations. In all these physical applications, one is interested in the thermodynamic limit where the size of the system tends to infinity and where details of the lattice become irrelevant. In the most basic scenario, one considers the infra-red limit and recovers this way conformal invariance. On the other hand, combinatorics is the study of discrete structures in mathematics. Combinatorial properties on integrable models will thus be uncovered by taking a different point of view on them which considers them on finite lattices and emphasizes their discrete properties. The purpose of this text is to show that the same methods and concepts of quantum integrability lead to non-trivial combinatorial results. The latter may be of intrinsic mathematical interest, in some cases proving, reproving, extending statements found in the literature. They may also lead back to physics by taking appropriate scaling limits. Let us be more specific on the kind of applications we have in mind. First and foremost comes the connection to enumerative combinatorics. For us the story begins in 1996, when Kuperberg showed how to enumerate alternating sign matrices using Izergin’s formula for the six-vertex model. The observation that alternating sign matrices are nothing but configurations of the six-vertex model in disguise, paved the way to a fruitful interaction between two subjects which were disjoint until then: (i) the study of alternating sign matrices, which began in the early eighties after their definition by Mills, Robbins, and Rumsey in relation to Dodgson condensation, and whose enumerative properties were studied in the following years, displaying remarkable connections with a much older class of combinatorial objects, namely plane partitions; and (ii) the study of the six-vertex model, one of the most fundamental solvable statistical models in two-dimensions, which was undertaken in the sixties and has remained at the center of the activity around quantum integrable models ever since. One of the most noteworthy recent chapters in this continuing story is the Razumov–Stroganov conjecture, in 2001, which emerged out of a collective effort by combinatorialists and physicists to understand the connection between the aforementioned objects and another class of statistical models, namely loop models. The work of the author was mostly a byproduct of various attempts to understand (and possibly prove) this conjecture. A large part of this manuscript is dedicated to reviewing these questions. Another interesting, related application is to algebraic combinatorics, due to the appearance of certain families of polynomials in quantum integrable models. In the context of the Razumov– Stroganov, they were introduced by Di Francesco and Zinn-Justin in 2004, but their true meaning was only clarified subsequently by Pasquier, creating a connection to representation theory of affine Hecke algebras and to previously studied classes of polynomials such as Macdonald polynomials. These polynomials satisfy relations which are typically studied in algebraic combinatorics, e.g. involving divided difference operators. The use of specific bases of spaces of polynomials, which is necessary for a combinatorial interpretation, connects to the theory of canonical bases and the work of Kazhdan and Lusztig. Finally, an exciting and fairly new aspect in this study of integrable models is to try to find an algebro-geometric interpretation of some of the objects and of the relations that satisfy. The most naive version of it would be to relate the integer numbers that appear in our models to problems in enumerative geometry, so that they become intersection numbers for certain algebraic varieties. A more sophisticated version involves equivariant cohomology or K-theory, which typically leads to polynomials instead of integers. The connection between integrable models and certain classes of polynomials with geometric meaning is not entirely new, and the work that will be described here bears some resemblance, as will be reminded here, to that of Fomin and Kirillov on Schubert and Grothendieck polynomials. However there are also novelties, including the use of the multidegree 4 technology of Knutson et al, and we apply these ideas to a broad class of models, resulting in new formulas for known algebraic varieties such as orbital varieties and the commuting variety, as well as in the discovery of new geometric objects, such as the Brauer loop scheme. The presentation that follows, though based on the articles of the author, is meant to be essentially self-contained. It is aimed at researchers and graduate students in mathematical physics or in combinatorics with an interest in exactly solvable statistical models. For simplicity, the inted with grable models that are defined are based on the underlying affine quantum group Uq (sl(2)), the notable exception of the discussion of the Brauer loop model in the last section. Furthermore, only the spin 1/2 representation and periodic boundary conditions are considered. There are interesting generalizations to higher rank, higher spin and to other boundary conditions of some of these results, on which the author has worked, but for these the reader is referred to the literature. The plan of this manuscript is the following. In section 1, we discuss free fermionic methods. Though free fermions in two dimensions may seem like an excessively simple physical model, they already provide a wealth of combinatorial formulae. In fact they have become extremely popular in the recent mathematical literature. We shall apply the basic formalism of free fermions to introduce Schur functions, and then spend some time reviewing the properties of the latter, because they will reappear many times in our discussion. We shall then briefly discuss the application to the enumeration of plane partitions. Section 2 covers the six-vertex model, and in particular the six-vertex model with domain wall boundary conditions. We shall discuss its quantum integrability, which is the root of its exact solvability. Then we shall apply it to the enumeration of alternating sign matrices. In section 3, we shall discuss statistical models of loops, their interrelations with the six-vertex model, their combinatorial properties and formulate the Razumov–Stroganov conjecture. Section 4 introduces the quantum Knizhnik–Zamolodchikov equation, which will be used to reconnect some of the objects discussed previously. The last section, 5, will be devoted to a brief review of the current status on the geometric reinterpretation of some of the concepts above, focusing on the central role of the quantum Knizhnik–Zamolodchikov equation. 1. Free fermionic methods As mentioned above, we want to spend some time defining a typical free fermionic model and to apply it to rederive some useful formulae for Schur functions, which will be needed later. We shall also need some formulae concerning the enumeration of plane partitions, which will appear at the end of this section. 1.1. Definitions. 1.1.1. Operators and Fock space. Consider a fermionic operator ψ(z): X X 1 1 ψ ⋆ (z) = ψ−k z k− 2 , ψk⋆ z k− 2 (1.1) ψ(z) = k∈Z+ 21 k∈Z+ 12 with anti-commutation relations (1.2) [ψr⋆ , ψs ]+ = δrs [ψr , ψs ]+ = [ψr⋆ , ψs⋆ ]+ = 0 ψ(z) and ψ ⋆ (z) should be thought of as generating series for the ψk and ψk⋆ , so that z is just a formal variable. What we have here is a complex (charged) fermion, with particles, and antiparticles which can be identified with holes in the Dirac sea. These fermions are one-dimensional, in the sense that their states are indexed by (half-odd-)integers; ψk⋆ creates a particle (or destroys a hole) at location k, whereas ψk destroys a particle (creates a hole) at location k. 5 We shall explicitly build the Fock space F and the representation of the fermionic operators now. Start from a vacuum |0i which satisfies (1.3) ψk⋆ |0i = 0 ψk |0i = 0 k > 0, that is, it is a Dirac sea filled up to location 0: |0i = · · · t t t t t td td td k<0 td td · · · 0 Then any state can be built by action of the ψk and ψk⋆ from |0i. In particular one can define more general vacua at level ℓ ∈ Z: ( ⋆ ⋆ ⋆ ℓ>0 ψℓ− 1 ψℓ− 3 · · · ψ 1 |0i 2 2 2 = · · · t t t t t td td td td td · · · (1.4) |ℓi = ψℓ+ 1 ψℓ+ 3 · · · ψ− 1 |0i ℓ < 0 2 2 2 ℓ which will be useful in what follows. They satisfy (1.5) ψk |ℓi = 0 ψk⋆ |ℓi = 0 k < ℓ k > ℓ, More generally, define a partition to be a weakly decreasing finite sequence of non-negative integers: λ1 ≥ λ2 ≥ · · · ≥ λn ≥ 0. We usually represent partitions as Young diagrams (also called Ferrers diagrams): for example λ = (5, 2, 1, 1) is depicted as λ= To each partition λ = (λ1 , . . . , λn ) we associate the following state in Fℓ : (1.6) ⋆ ⋆ ⋆ · · · ψℓ+λ |λ; ℓi = ψℓ+λ 1 |ℓ − ni 1ψ ℓ+λ2 − 3 n −n+ 1− 2 2 2 Note the important property that if one “pads” a partition with extra zeroes, then the corresponding state remains unchanged. In particular for the empty diagram ∅, |∅; ℓi = |ℓi. For ℓ = 0 we just write |λ; 0i = |λi. This definition has the following nice graphical interpretation: the state |λ; ℓi can be described by numbering the edges of the boundary of the Young diagram, in such a way that the main diagonal passes between ℓ − 12 and ℓ + 21 ; then the occupied (resp. empty) sites correspond to vertical (resp. horizontal) edges. With the example above and ℓ = 0, we find (only the occupied sites are numbered for clarity) dt dt dt . . . dt t12 dt dt dt t92 t−32 dt t−52 t−29 t−11 2 .. . The |λ; ℓi, where λ runs over all possible partitions (two partitions being identified if they are obtained from each other by adding or removing zero parts), form an orthonormal basis of a subspace of F which we denote by Fℓ . ψk and ψk⋆ are Hermitean conjugate of each other. 6 Note that (1.6) fixes our sign convention of the states. In particular, this implies that when one acts with ψk (resp. ψk⋆ ) on a state |λi with a particle (resp. a hole) at k, one produces a new state |λ′ i with the particle removed (resp. added) at k times −1 to the power the number of particles to the right of k. The states λ can also be produced from the vacuum by acting with ψ to create holes; paying attention to the sign issue, we find |λ; ℓi = (−1)|λ| ψℓ−λ′ + 1 · · · ψℓ−λ′m +m− 1 |ℓ + mi (1.7) 1 2 2 λ′i where the are the lengths of the columns of λ, |λ| is the number of boxes of λ and m = λ1 . This formula is formally identical to (1.6) if we renumber the states from right to left, exchange ψ and ψ ⋆ , and replace λ with its transpose diagram λ′ (this property is graphically clear). So the particle–hole duality translates into transposition of Young diagrams. Finally, introduce the normal ordering with respect to the vacuum |0i: ( ψj⋆ ψk j>0 ⋆ ⋆ (1.8) : ψj ψk : = − : ψk ψj : = ⋆ −ψk ψj j < 0 which allows to get rid of trivial infinite quantities. [ action. The operators ψ ⋆ (z)ψ(w) give rise to the Schwinger representation 1.1.2. gl(∞) and gl(1) of gl(∞) on F, whose usual basis is the : ψr⋆ ψs : , r, s ∈ Z + 21 , and the identity. In the first quantized picture this representation is simply the natural action of gl(∞) on the one-particle Hilbert space P 1 CZ+ 2 and exterior products thereof. The electric charge J0 = r : ψr⋆ ψr : is a conserved number and classifies the irreducible representations of gl(∞) inside F, which are all isomorphic. The highest weight vectors are precisely our vacua |ℓi, ℓ ∈ Z, so that F = ⊕ℓ∈Z Fℓ with Fℓ the subspace in which J0 = ℓ. The gl(1) current j(z) = : ψ ⋆ (z)ψ(z) : = (1.9) X Jn z −n−1 n∈Z with Jn = (1.10) P r ⋆ ψ : forms a gl(1) [ (Heisenberg) sub-algebra of gl(∞): : ψr−n r [Jm , Jn ] = mδm,−n Note that positive modes commute among themselves. This allows to define the general “Hamiltonian” ∞ X tq Jq (1.11) H[t] = q=1 where t = (t1 , . . . , tq , . . .) is a set of parameters (“times”). The Jq , q > 0, displace one of the fermions q steps to the left. This is expressed by the formulae describing the time evolution of the fermionic fields: (1.12) eH[t] ψ(z)e−H[t] = e− P∞ + P∞ H[t] e ⋆ −H[t] ψ (z)e =e q=1 tq z q q q=1 tq z ψ(z) ψ ⋆ (z) (proof: compute [Jq , ψ [⋆] (z)] = ±z q ψ [⋆] (z) and exponentiate). Of course, similarly, J−q , q > 0, moves one fermion q steps to the right. 7 1.2. Schur functions. 1.2.1. Free fermionic definition. It is known that the map |Φi 7→ hℓ| eH[t] |Φi is an isomorphism from Fℓ to the space of polynomials in an infinite number of variables t1 , . . . , tq , . . .. Thus, we obtain a basis of the latter as follows: for a given Young diagram λ, define the Schur function sλ [t] by sλ [t] = hℓ| eH[t] |λ; ℓi (1.13) (by translational invariance it is in fact independent of ℓ). In the language of Schur functions, the tq are (up to a conventional factor 1/q) the power sums, see section 1.2.3 below. We provide here various expressions of sλ [t] using the free fermionic formalism. In fact, many of the methods used are equally applicable to the following more general quantity: sλ/µ [t] = hµ; ℓ| eH[t] |λ; ℓi (1.14) where λ and µ are two partitions. It is easy to see that in order for sλ/µ [t] to be non-zero, µ ⊂ λ as Young diagrams; in this case sλ/µ is known as the skew Schur function associated to the skew Young diagram λ/µ. The latter is depicted as the complement of µ inside λ. This is appropriate because skew Schur functions factorize in terms of the connected components of the skew Young diagram λ/µ. Examples: s = t1 , s = 12 t21 − t2 , s s = s2 = t21 , s = 5 4 24 t1 = 12 t21 + t2 , s = 13 t31 − t3 . + 21 t21 t2 + 12 t22 − t1 t3 − t4 . 1.2.2. Wick theorem and Jacobi–Trudi identity. First, we apply the Wick theorem. Consider as the definition of the time evolution of fermionic fields: ψk [t] = eH[t] ψk e−H[t] (1.15) ψk⋆ [t] = eH[t] ψk⋆ e−H[t] In fact, (1.12) gives us the “solution” of the equations of motion in terms of the generating series ψ(z), ψ ⋆ (z). Noting that the Hamiltonian is quadratic in the fields, we now state the Wick theorem: (1.16) hℓ| ψi1 [0] · · · ψin [0]ψj⋆1 [t] . . . ψj⋆n [t] |ℓi = det hℓ| ψip [0]ψj⋆q [t] |ℓi 1≤p,q≤n Next, start from the expression (1.14) of sλ/µ [t]: padding with zeroes λ or µ so that they have the same number of parts n, we can write sλ/µ [t] = h−n| ψµn −n+ 1 · · · ψµ1 − 1 eH[t] ψλ⋆ 1 − 1 · · · ψλ⋆ n −n+ 1 |−ni 2 2 2 2 and apply the Wick theorem to find: sλ/µ [t] = It is easy to see that us denote it (1.17) det h−n| ψµp −p+ 1 eH[t] ψλ⋆ q −q+ 1 |−ni 1≤p,q≤n h−n| ψi eH[t] ψj⋆ 2 2 |−ni does not depend on n and thus only depends on j − i. Let ⋆ hk [t] = h1| eH[t] ψk+ 1 |0i 2 X k≥0 P t zq H[t] ⋆ hk [t]z = h1| e ψ (z) |0i = e q≥1 q k (k = j − i; note that hk [t] = 0 for k < 0). 8 The final formula we obtain is (1.18) sλ/µ [t] = hλq −µp −q+p [t] det 1≤p,q≤n or, for regular Schur functions, (1.19) sλ [t] = det 1≤p,q≤n This is known as the Jacobi–Trudi identity. hλq −q+p [t] By using “particle–hole duality”, we can find a dual form of this identity. We describe our states in terms of hole positions, parametrized by the lengths of the columns λ′p and µ′q , according to (1.7): s [t] = (−1)|λ|+|µ| hm| ψ ⋆ · · · ψ⋆ eH[t] ψ ′ 1 · · · ψ 1 |mi λ/µ −µ′m +m− 21 −µ′1 + 12 −λ′m +m− 2 −λ1 + 2 Again the Wick theorem applies and expresses sλ/µ in terms of the two point-function hm| ψi⋆ eH[t] ψj |mi, which only depends on i − j = k and is given by (1.20) P X (−1)q−1 tq z q k H[t] k H[t] e [t] = (−1) h−1| e ψ e [t]z = h−1| e ψ(−z) |0i = e q≥1 1 |0i k k −k+ 2 k≥0 The finally formula takes the form (1.21) sλ/µ [t] = det 1≤p,q≤n or, for regular Schur functions, (1.22) sλ [t] = det eλ′q −µ′p −q+p [t] 1≤p,q≤n eλ′q −q+p [t] This is the dual Jacobi–Trudi identity, also known as Von Nägelsbach–Kostka identity. 1.2.3. Weyl formula. In the following sections 1.2.3–1.2.6, we shall fix an integer n and P consider the following change of variable (this is essentially the Miwa transformation [79]) tq = 1q ni=1 xqi . The Schur function becomes a symmetric polynomial of these variables xi , which we denote by sλ (x1 , . . . , xn ), and we now derive a different (first quantized) formula for it. Due to obvious translational invariance of all the operators involved, we may as well set ℓ = n. Use the definition (1.6) of |λi and the commutation relations (1.12) to rewrite the left hand side as P Pn t zq hn| eH[t] |λ; ni = e q≥1 q i=1 i hn| ψ ⋆ (z1 )ψ ⋆ (z2 ) · · · ψ ⋆ (zn ) |0i z n+λ1 −1 z n+λ2 −2 ...z λn n 1 2 where ... means picking one term in a generating series. We can easily evaluate the remaining bra-ket to be: (we now use the ℓ = 0 notation for the l.h.s.) P Pn q Y t z (zi − zj )z n+λ1 −1 z n+λ2−2 ...z λn h0| eH[t] |λi = e q≥1 q i=1 i 1≤i<j≤n Pn P q i=1 zi q≥1 tq P Now write tq = 1q nj=1 xqj and note that e of) the Cauchy determinant: 1 = Qn i,j=1 (1 2 − zi xj )−1 . We recognize (part det1≤i,j≤n (1 − xi zj )−1 Q h0| eH[t] |λi = n+λ −1 n+λ −2 λn z1 1 z2 2 ···zn i<j (xi − xj ) 9 n At this stage we can just expand separately each column of the matrix (1 − xi zj )−1 to pick the right power of zj ; we find: λ +n−j det1≤i,j≤n (xi j ) H[t] Q h0| e |λi = (x − x ) j i<j i (1.23) Remark 1: defined in terms of a fixed number n of variables, as in (1.23), sλ (x1 , . . . , xn ) has the following group-theoretic interpretation. The polynomial irreducible representations of GL(n) are known to be indexed by partitions. Then sλ (x1 , . . . , xn ) is the character of representation λ evaluated at the diagonal matrix Qdiag(x1 , . . . , xn ). Hence, the dimension of λ as a GL(n) representation is given by sλ (1, . . . , 1) = 1≤i<j≤n (λi − i − λj + j)/(j − i). | {z } n P Remark 2: the more general Miwa transformation allows for coefficients: tq = 1q ni=1 αi xqi . In particular if we use minus signs, we get the notion of plethystic negation. Combining it with the usual negation of variables is equivalent to transposing Young diagrams: indeed it amounts to exchanging the hk [t] and the ek [t]. In other words, sλ [t] = sλ′ [−ǫt] − ǫtq := (−1)q−1 tq More generally, one defines Schur function sλ (x1 , . . . , xn /y1 , . . . , ym ) to be equal P the supersymmetric P q ). (−y ) to sλ [t] where tq = 1q ( ni=1 xqi − m i i=1 1.2.4. Schur functions and lattice fermions. Note that the change of variables tq = us to write n Y X xq H[t] eφ+ (xi ) φ+ (x) = e = Jq q i=1 1 q Pn q j=1 xj allows q≥1 So we can think of the “time evolution” as a series of discrete steps represented by commuting operators exp φ+ (xi ). In the language of statistical mechanics, these are transfer matrices (and the existence of a one-parameter family of commuting transfer matrices exp φ+ (x) is of course related to the integrability of the model). We now show that they have a very simple meaning in terms of lattice fermions. Consider a two-dimensional square lattice, one direction being our space Z + 12 and one direction being time. In what follows we shall reverse the arrow of time (that is, we shall consider that time flows upwards on the pictures), which makes the discussion slightly easier since products of operators are read from left to right. The rule to go from one step to the next according to the evolution operator exp φ+ (x) can be formulated either in terms of particles or in terms of holes: • Each particle can go straight or hop to the right as long as it does not reach the (original) location of the next particle. Each step to the right is given a weight of x. • Each hole can only go straight or one step to the left as long as it does not bump into its neighbor. Each step to the left is given a weight of x. Obviously the second description is simpler. An example of a possible evolution of the system with given initial and final states is shown on Fig. 1(a). The proof of these rules consists in computing explicitly hµ| eφ+ (x) |λi by applying say (1.21) for tq = 1q xq , and noting that in this case, according to (1.20), en [t] = 0 for n > 1. This strongly constrains the possible transitions and produces the description above. 10 h x x h x h x x h x x h x x h x x h x h x x h x x h x x h x x h x h x x x h x h x x h x x h x x x h x h x h x 1 1 3 2 4 1 (a) (b) Figure 1. A lattice fermion configuration and the corresponding (skew) SSYT. 1.2.5. Relation to Semi-Standard Young tableaux. A semi-standard Young tableau (SSYT) of shape λ is a filling of the Young diagram of λ with elements of some ordered alphabet, in such a way that rows are weakly increasing and columns are strictly increasing. We shall use here the alphabet {1, 2, . . . , n}. For example with λ = (5, 2, 1, 1) one possible SSYT with n ≥ 5 is: 1 2 4 5 5 3 3 4 5 It is useful to think of Young tableaux as time-dependent Young diagrams where the number indicates the step at which a given box was created. Thus, with the same example, we get ∅, , , , , =λ So a Young tableau is nothing but a statistical configuration of our lattice fermions, where the initial state is the vacuum. Similarly, a skew SSYT is a filling of a skew Young diagram with the same rules; it corresponds to a statistical configuration of lattice fermions with arbitrary initial and final states. The correspondence is exemplified on Fig. 1(b). Each extra box corresponds to a step to the right for particles or to the left for holes. The initial and final states are ∅ and λ, which is the case for Schur functions, cf (1.13). We conclude that the following formula holds: X Y (1.24) sλ (x1 , . . . , xn ) = xTb T ∈SSYT(λ,n) b box of T This is often taken as a definition of Schur functions. It is explicitly stable with respect to n in the sense that sλ (x1 , . . . , xn , 0, . . . , 0) = sλ (x1 , . . . , xn ). It is however not obvious from it that sλ is symmetric by permutation of its variables. This fact is a manifestation of the underlying free fermionic (“integrable”) behavior. Of course an identical formula holds for the more general case of skew Schur functions. 11 x x > ∧ ∧ ∧ > > > x x > ∧ ∧ ∧ x > > > x > ∧ ∧ ∧ x > > > x > ∧ ∧ ∧ x > > > x > ∧ ∧ ∧ x > > > x > ∧ ∧ ∧ > > > x ∧ ∧ ∧ x x h x h x h x h x h x h x h ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ ∧ x h x h x h x h x h x h x h Figure 2. Underlying directed graphs for particles and holes. 1.2.6. Non-Intersecting Lattice Paths and Lindström–Gessel–Viennot formula. The rules of evolution given in section 1.2.4 strongly suggest the following explicit description of the lattice fermion configurations. Consider the directed graphs of Fig. 2 (the graphs are in principle infinite to the left and right, but any given bra-ket evaluation only involves a finite number of particles and holes and therefore the graphs can be truncated to a finite part). Consider Non-Intersecting Lattice Paths (NILPs) on these graphs: they are paths with given starting points (at the bottom) and given ending points (at the top), which follow the edges of the graph respecting the orientation of the arrows, and which are not allowed to touch at any vertices. One can check that the trajectories of holes and particles following the rules described in section 1.2.4 are exactly the most general NILPs on these graphs. In this context, the Jacobi–Trudi identity (1.19) becomes a consequence of the so-called Lindström–Gessel–Viennot formula [72, 36]. This formula expresses N (i1 , . . . , in ; j1 , . . . , jn ), the weighted sum of NILPs on a general directed acyclic graph from starting locations i1 , . . . , in to ending locations j1 , . . . , jn , where the weight of a path is the products of weights of the edges, as (1.25) N (i1 , . . . , in ; j1 , . . . , jn ) = det N (ip ; jq ) p,q More precisely, in Lindström’s formula, sets of NILPs such that the path starting from ik ends at jw(k) get an extra sign which is that of the permutation w. This is nothing but the Wick theorem once again (but with fermions living on a general graph), and from this point of view is a simple exercise in Grassmannian Gaussian integrals. In the special case of a planar graph with appropriate starting points (no paths are possible between them) and ending points, only one permutation, say the identity up to relabelling, contributes. In order to use this formula, one only needs to compute N (i; j), the weighted sum of paths from i to j. Let us do so in our problem. In the case of particles (left graph), numbering the initial and final points from left to right, we find that the weighted sum of paths from i to j, where a weight xi is given to each right move at time-step i, only depends on j − i; if we denote it by hj−i (x1 , . . . , xn ), we have the obvious generating series formula n Y X 1 hk (x1 , . . . , xn )z k = 1 − zxi i=1 k≥0 Note that this formula coincides with the alternate definition (1.17) of hk [t] if we set as usual 1 Pn tq = q i=1 xqi . Thus, applying the LGV formula (1.25) and choosing the correct initial and final points for Schur functions or skew Schur functions, we recover immediately (1.18,1.19). 12 In the case of holes (right graph), numbering the initial and final points from right to left, we find once again that the weighted sum of paths from i to j, where a weight xi is given to each left move at time-step i, only depends on j − i; if we denote it by ej−i (x1 , . . . , xn ), we have the equally obvious generating series formula X k ek (x1 , . . . , xn )z = n Y (1 + zxi ) i=1 k≥0 which coincides with (1.20), thus allowing us to recover (1.21,1.22). 1.2.7. Relation to Standard Young Tableaux. A Standard Young Tableau (SYT) of shape λ is a filling of the Young diagram of λ with elements of some ordered alphabet, in such a way that both rows and columns are strictly increasing. There is no loss of generality in assuming that the alphabet is {1, . . . , n}, where n = |λ| is the number of boxes of λ. For example, 1 2 6 8 9 3 4 5 7 is a SYT of shape (5, 2, 1, 1). Standard Young Tableaux are connected to the representation theory of the symmetric group; the number of such tableaux with given shape λ is the dimension of λ as an irreducible representation of the symmetric group, which is up to a factor n! the evaluation of the Schur function sλ at tq = δ1q . Indeed, in this case one has H[t] = J1 , and there is only one term contributing to the bra-ket hλ| eH[t] |0i in the expansion of the exponential: 1 hλ| J1n |0i n! In terms of lattice fermions, J1 has a direct interpretation as the transfer matrix for one particle hopping one step to the left. As the notion of SYT is invariant by transposition, particles and holes play a symmetric role so that the evolution can be summarized by either of the two rules: sλ [δ1· ] = • Exactly one particle moves one step to the right in such a way that it does not bump into its neighbor; all the other particles go straight. • Exactly one hole moves one step to the left in such a way that it does not bump into its neighbor; all the other holes go straight. An example of such a configuration is given on Fig. 3. ⋆ 1.2.8. Cauchy formula. As an additional remark, consider the commutation of eH[t] and eH [u] , where H ⋆ [u], the transpose of H[u], is obtained from it by replacing Jq with J−q . Using the Baker–Campbell–Hausdorff formula and the commutation relations (1.10) we find P ⋆ ⋆ [u] qt u H[t] H e e = e q≥1 q q eH [u] eH[t] P q 1 eφ− (y) eφ+ (x) with φ± (x) = q≥1 xq J±q . or equivalently eφ+ (x) eφ− (y) = 1−xy If we now use the fact that the |λi form a basis of F0 , we obtain the Cauchy formula: P X Y ⋆ qt u (1.26) h0| eH[t] eH [u] |0i = sλ [t]sλ [u] = (1 − xi yj )−1 = e q≥1 q q i,j λ 13 x h x h x x x h x x h x x h x x h x x h x x x h x h x x x h x x h x h x x x x h x h x h x 1 3 2 4 ∅ (a) (b) Figure 3. A lattice fermion configuration and the corresponding SYT. (b) (a) Figure 4. (a) A plane partition of size 2 × 3 × 4. (b) The corresponding dimer configuration. with tq = 1 q Pn q i=1 xi , uq = 1 q Pn q i=1 yi . 1.3. Application: Plane Partition enumeration. Plane partitions are a well-known class of combinatorial objects. The name originates from the way they were first introduced [74] as twodimensional generalizations of partitions; here we shall directly define plane partitions graphically. Their study has a long history in mathematics, with a renewal of interest in the eighties [97] in combinatorics, and more recently in mathematical physics [86]. 1.3.1. Definition. Intuitively, plane partitions are pilings of boxes (cubes) in the corner of a room, subject to the constraints of gravity. An example is given on Fig. 4(a). Typically, we ask for the cubes to be contained inside a bigger box (parallelepiped) of given sizes. Alternatively, one can project the picture onto a two-dimensional plane (which is inevitably what we do when we draw the picture on paper) and the result is a tiling of a region of the plane by lozenges (rhombi with 60/120 degrees angles) of three possible orientations, as shown on the right of the figure. If the cubes are inside a parallelepiped of size a × b × c, then, possibly drawing the walls of the room as extra tiles, we obtain a lozenge tiling of a hexagon with sides a, b, c, which is the situation we consider now. Note that each lozenge is the union of two adjacent triangles which live on an underlying fixed triangular lattice. So this is a statistical model on a regular lattice. In fact, we can identify it with a model of dimers living on the dual lattice, that is the honeycomb lattice. Each lozenge corresponds to an occupied edge, see Fig. 4(b). Dimer models have a long history of their own (most notably, 14 Kasteleyn’s formula [51] is the standard route to their exact solution, which we do not use here), which we cannot possibly review here. 1.3.2. MacMahon formula. In order to display the free fermionic nature of plane partitions, we shall consider the following operation. In the 3D view, consider slices of the piling of boxes by hyperplanes parallel to two of the three axis and such that they are located half-way between successive rows of cubes. In the 2D view, this corresponds to selecting two orientations among the three orientations of the lozenges and building paths out of these. Fig. 5 shows on the left the result of such an operation: a set of lines going from one side to the opposite side of the hexagon. They are by definition non-intersecting and can only move in two directions. Inversely, any set of such NILPs produces a plane partition. At this stage one can apply the LGV formula. But there is no need since this is actually the case already considered in section 1.3.4. Compare Figs. 5 and 1: the trajectories of holes are exactly our paths (the trajectories of particles form another set of NILPs corresponding to another choice of two orientations of lozenges). If we attach a weight of xi to each blue lozenge at step i, we find that the weighted enumeration of plane partitions in a a × b × c box is given by: Na,b,c (x1 , . . . , xa+b ) = h0| eH[t] |b × ci = sb×c (x1 , . . . , xa+b ) where b×c is the rectangular Young diagram with height b and width c. In particular the unweighted enumeration is the dimension of the Young diagram b × c as a GL(a + b) representation: (1.27) Na,b,c = c b Y a Y Y i+j +k−1 i+j +k−2 i=1 j=1 k=1 which is the celebrated MacMahon formula. But the more general formula provides various refinements. For example, one can assign a weight of q to each cube in the 3D picture. It can be shown that this is achieved by setting xi = q a+b−i (up to a global power of q). This way we find the q-deformed formula c b Y a Y Y 1 − q i+j+k−1 Na,b,c (q) = 1 − q i+j+k−2 i=1 j=1 k=1 Many more formulae can be obtained in this formalism. The reader may for example prove that X Na,b,c = sλ (1, . . . , 1)sλ (1, . . . , 1) | {z } | {z } λ:λ1 ≤c a b or that Na,b,c = det(1 + Tc×b Tb×a Ta×c ) (where Ty×x is the matrix with y rows and x columns and entries ji , i = 0, . . . , y − 1, j = 0, . . . , x − 1), as well as investigate their possible refinements. (for more formulae similar to the last one, see [30]). Finally, one can take the limit a, b, c → ∞, and by comparing the power of the factors 1 − q a in the numerator and the denominator, one finds another classical formula ∞ Y (1 − q n )−n N∞,∞,∞(q) = n=1 Note that our description in terms of paths clearly breaks the threefold symmetry of the original hexagon. It strongly suggests that one should be able to introduce three series of parameters to provide an even more refined counting of plane partitions. With two sets of parameters, this is in fact known in the combinatorial literature and is related to so-called double Schur functions 15 a b c Figure 5. NILPs corresponding to a plane partition. (these will reappear in section 5.2.5). The full three-parameter generalization is less well-known and appears in [107], as will be recalled in section 4.3.2. Remark: as the name suggests, plane partitions are higher dimensional versions of partitions, that is of Young diagrams. After all, each slice we have used to define our NILPs is also a Young diagram itself. However these Young diagrams should not be confused with the ones obtained from the NILPs by the correspondence of section 1.2. 1.3.3. Totally Symmetric Self-Complementary Plane Partitions. In the mathematical literature, many more complicated enumeration problems are addressed, see [97]. In particular, consider lozenge tilings of a hexagon of shape 2a×2a×2a. One notes that there is a group of transformations acting naturally on the set of configurations. We consider here the dihedral group of order 12 which is consists of rotations of π/3 and reflections w.r.t. axis going through opposite corners of the hexagon or through middles of opposite edges. To each of its subgroups one can associate an enumeration problem. Here we discuss only the case of maximal symmetry, i.e. the enumeration of Plane Partitions with the dihedral symmetry. They are called in this case Totally Symmetric Self-Complementary Plane Partitions (TSSCPPs). The fundamental domain is a twelfth of the hexagon, see Fig. 6. Inside this fundamental domain, one can use the equivalence to NILPs by considering green and blue lozenges. However it is clear that the resulting NILPs are not of the same type as those considered before for general plane partitions, for two reasons: (i) the starting and ending points are not on parallel lines, and (ii) the endpoints are in fact free to lie anywhere on a vertical line. However the LGV formula still holds. For future purposes we provide an integral formula for the counting of TSSCPPs where a weight τ is attached to every blue lozenge in the fundamental domain [28]. Let us call rj the location of the endpoint of the j th path, numbered from top to bottom starting at zero. We first apply the LGV formula to write the numberi of NILPs with given endpoints to be i det(Ni,rj )1≤i,j≤n−1 where Ni,r = τ 2i−r−1 2i−r−1 = (1 + τ u) |u2i−r−1 . Next we sum over them and 16 Figure 6. A TSSCPP and the associated NILP. obtain Nn (τ ) = X 0≤r1 <r2 <···<rn−1 (1.28) = n−1 Y (1 + τ ui )i i=1 r det[(1 + τ ui )i ui j ]Qn−1 u2i−1 i=1 X 0≤r1 <r2 <···<rn−1 i r det(ui j )Qn−1 u2i−1 i=1 i We recognize the numerator of a Schur function; the summation isPsimply over all YoungQ diagrams n−1 (1 − with nQparts. At this stage we use a classical summation formula, λ sλ (u1 , . . . , un−1 ) = i=1 −1 −1 ui ) 1≤i<j≤n−1 (1 − ui uj ) , to conclude that (1.29) Nn (τ ) = Y 1≤i<j≤n−1 n−1 uj − ui Y (1 + τ ui )i Qn−1 2i−1 i=1 ui 1 − ui uj 1 − ui i=1 where Qn−1 u2i−1 is now interpreted as picking the coefficient of a monomial in a powers series i=1 i around zero. This formula can be used to generate efficiently these polynomials by computer; in particular, we find the numbers Nn (1) = 1, 2, 7, 42, 429 . . . which have only small prime factors. This allows to conjecture a simple product form: Nn (1) = n−1 Y i=0 (3i + 1)! 1!4! . . . (3n − 2)! = (n + i)! n!(n + 1)! . . . (2n − 1)! which was in fact proven in [2]. As a byproduct of what follows (sections 2.5 and 4.4), we shall obtain a (rather indirect) derivation of this evaluation. 1.4. Classical integrability. The free fermionic Fock space is also important for the construction of solutions of classically integrable hierarchies. We cannot possibly describe these important ideas here, and refer the reader to [44] and references therein for details. Since an explicit example will appear in section 2, let us simply say a few general words. Recall the isomorphism Φ 7→ hℓ| eH[t] |Φi from Fℓ to the space of polynomials in the variables tq (or equivalently to the space of symmetric 17 Figure 7. All TSSCPPs of size 1, 2, 3. Figure 8. A configuration of the six-vertex model. functions if the tq are interpreted as power sums). The resulting function will be a tau-function of the Kadomtsev–Petiashvili (KP) hierarchy (as a function of the tq ) for appropriately chosen |Φi. By appropriately chosen we mean the following. In the first quantized picture, the essential property of free fermions is the possibility to write their wave function as a Slater determinant; this amounts to considering states which are exterior products of one-particle states. Geometrically this is interpreted as saying that the state (defined up to multiplication by a scalar) really lives in a subspace of the full Hilbert space called a Grassmannian. The equations defining this space (Plücker relations) are quadratic; these equations are differential equations satisfied by hℓ| eH[t] |Φi. They are Hirota’s form of the equations defining the KP hierarchy. In section 2 we shall find ourselves in a slightly more elaborate setting, which results in the Toda lattice hierarchy. 2. The six-vertex model The six vertex model is an important model of classical statistical mechanics in two dimensions, being the prototypical (vertex) integrable model. The ice model (infinite temperature limit of the six-vertex model) was solved by Lieb [69] in 1967 by means of Bethe Ansatz, followed by several generalizations [68, 70, 71]. The solution of the most general six vertex model was given by Sutherland [99] in 1967. The bulk free energy was calculated in these papers for periodic boundary conditions (PBC). Here our main interest will be in a different kind of boundary conditions, the so-called Domain Wall Boundary Conditions. But first we provide a brief review of the six-vertex model. 18 a1 a2 b1 b2 c1 c2 Figure 9. Weights of the six-vertex model. 2.1. Definition. 2.1.1. Configurations. The six-vertex model is defined on a (subset of the) square lattice by putting arrows (two possible directions) on each edge of the lattice, with the additional rule that at each vertex, there are as many incoming arrows as outgoing ones. See Fig. 8 for an example, and for two alternative descriptions: the “square ice” version in which arrows represent which oxygen atom (sitting at each lattice vertex) the hydrogen ions (living on the edges) are closer to, with the “ice rule” that exactly two hydrogen ions are close to each oxygen atom; and the “path” version in which one considers edges with right or up arrows as occupied, so that they form north-east going paths. Around a given vertex, there are only 6 configurations of edges which respect the arrow conservation rule, see Fig. 9, hence the name of the model. 2.1.2. Weights. The weights are assigned to the six vertices, see Fig. 9. Thus the partition function is given by X Y Z= (weight of the vertex) configurations vertex An additional remark is useful. With any fixed boundary conditions, one can show that the difference between the numbers of vertices of the two types c is constant (independent of the configuration). This means that only the product c2 = c1 c2 of their two weights matters. Let us denote similarly a2 = a1 a2 and b2 = b1 b2 . One can write a1 = ae+Ex +Ey b1 = be−Ex +Ey a2 = ae−Ex −Ey b2 = be+Ex −Ey and consider that a, b, c are the weights of the vertices, while Ex , Ey are electric fields. In what follows, we shall consider by default the model without any electric field, where the Boltzmann weights are invariant by reversal of every arrow and a1 = a2 = a, a1 = a2 = a, a1 = a2 = a; and sometimes comment on the generalization to non-zero fields. There is another way to formulate the partition function, using a transfer matrix. In order to set up a transfer matrix formalism, we first need to specify the boundary conditions. Let us consider doubly periodic boundary conditions in the two directions of the lattice, so that the model is defined on lattice of size M × L with the topology of a torus. Then one can write Z = tr TLM where TL is the 2L × 2L transfer matrix which corresponds to a periodic strip of size L. Explicitly, the indices of the matrix TL are sequences of L up/down arrows. TL can itself be expressed as a product of matrices which encode the vertex weights; in the case of integrable models, we usually 19 denote this matrix by the letter R: (2.1) Then we have (2.2) →↑ →↑ a →↓ 0 R= ←↑ 0 ←↓ 0 →↓ ←↑ ←↓ 0 0 0 b c 0 c b 0 0 0 a TL = tr0 (R0L · · · R02 R01 ) = · · · ··· 0 1 2 3 4 ith where Rij means the matrix R acting on the tensor product of and j th spaces, and 0 is an additional auxiliary space encoding the horizontal edges, as on the picture (note that the trace is on the auxiliary space and graphically means that the horizontal line reconnects with itself). On the picture “time” flows upwards and to the right. The introduction of a vertical electric field amounts to multiplying the transfer matrix by an z operator which commutes with it, of the form eEy Σ (Σz being the number of up arrows minus the number of down arrows). More interestingly, adding a horizontal field amounts to twisting the periodic transfer matrix: indeed, all the horizontal fields, using conservation of arrows at each vertex, can be moved to a single site, so that the transfer matrix becomes, up to conjugation by z eEx Σ , (2.3) TL = tr0 (R0L · · · R02 R01 Ω) z where the twist Ω acts on the auxiliary space and is of the form Ω = eLEx σ . 2.2. Integrability. 2.2.1. Properties of the R-matrix. Let us now introduce the following parametrization of the weights: a = q x − q −1 x−1 b = x − x−1 (2.4) c = q − q −1 x, q are enough to parametrize them up to global scaling. Instead of q one often uses q + q −1 a2 + b2 − c2 = 2ab 2 In general, q or ∆ are fixed whereas x is a variable parameter, called spectral parameter. It can be thought itself as a ratio of two spectral parameters attached to the lines crossing at the vertex. ∆= The matrix R(x) then satisfies the following remarkable identity: (Yang–Baxter equation) 1 R12 (x2 /x1 )R13 (x3 /x1 )R23 (x3 /x2 ) = R23 (x3 /x2 )R13 (x3 /x1 )R12 (x2 /x1 ) 1 2 3 2 3 This is formally the same equation that is satisfied by S matrices in an integrable field theory (field theory with factorized scattering, i.e. such that every S matrix is a product of two-body S matrices). 20 1 a/c ∆ = F ∆ ∆ = 1 D 1 = F − 1 AF ∆ = −∞ b/c 1 Figure 10. Phase diagram of the six-vertex model. The R-matrix also satisfies the unitarity equation: −1 R12 (x)R21 (x ) = (q x − q −1 −1 x )(q x −1 −q −1 1 1 2 2 x) with x = x2 /x1 . The scalar function could of course be absorbed by appropriate normalization of R. 2.2.2. Commuting transfer matrices. Consider now the transfer matrix as a function of the spectral parameter x, possibly with a twist: (2.5) TL (x) = tr0 (R0L (x) · · · R02 (x)R01 (x)Ω) Then using the Yang–Baxter equation repeatedly one obtains the relation [TL (x), TL (x′ )] = 0 We thus have an infinite family of commuting operators. In practice, for a finite chain TL (x) is a Laurent polynomial of x so there is a finite number of independent operators. Note that we could have used the more general inhomogeneous transfer matrix TL (x0 ; x1 , . . . , xL ) = tr0 (R0L (yL /x0 ) · · · R02 (y2 /x0 )R01 (y1 /x0 )Ω) where now we have spectral parameters yi attached to each vertical line i and one more parameter x0 attached to the auxiliary line. Then the same commutation relations hold for fixed yi and variable x0 . As is well-known, the commutation of the transfer matrices is only one relation in the Yang– Baxter algebra generated by the so-called RTT relations. The latter lead to an exact solution of the model using Algebraic Bethe Ansatz [31]. 2.3. Phase diagram. The phase diagram of the six-vertex model in the absence of electric field is discussed in great detail in chapter 8 of [4]. It can be deduced from the exact solution of the model using Bethe Ansatz after taking the thermodynamic limit. The physical properties of the system depend only on ∆ = (q + q −1 )/2, x playing the role of a lattice anisotropy parameter. We distinguish three phases, see Fig. 10: 21 (a) (b) or (c) or Figure 11. Correspondence between (a) (∆ = 0) six-vertex (b) NILPs and (c) domino tilings. (1) ∆ ≥ 1: the ferroelectric phase. This phase is non-critical. Furthermore, there are no local degrees of freedom: the system is frozen in regions filled with one of the vertices of type a or b (i.e. all arrows aligned), and no local changes (that respect arrow conservation) are possible. (2) ∆ < −1: the anti-ferroelectric phase. This phase is non-critical. This time there is a finite correlation length. The ground state of the transfer matrix corresponds to a state with zero polarization (in the limit ∆ → −∞, it is simply an alternation of up and down arrows). (3) −1 ≤ ∆ < 1: the disordered phase. This phase is critical. It possesses a continuum limit with conformal symmetry, and this limiting infra-red Conformal Field Theory is well-known: 1 , it is simply the c = 1 theory of a free boson on a circle with radius R given by R2 = 2(1−γ/π) ∆ = − cos γ, 0 < γ < π. The phase diagram in the presence of an electric field is more complicated, though the basic division into the three phases above remains. See [95, 84] for a description.1 2.4. Free fermion point. Inside the disordered phase, there is a special point ∆ = 0. We provide various representations of the six-vertex model which display the free fermionic behavior of this region of parameter space. 2.4.1. NILP representation. It is tempting to try to interpret the “north-east going paths” of Fig. 8 as Non-Intersecting Lattice Paths. The problem is that they can touch at vertices. One way to fix it is to consider the slightly modified paths of Fig. 11(b) The rule is to replace each vertex of (a) with the corresponding dotted square of (b) and then patch together the latter to form the paths.2 Note that the correspondence is no longer one-to-one: each vertex of type c1 corresponds to two possible local paths. 1Note that the discussion of the phase diagram in [104] is incomplete. 2Going from (a) to (b) amounts to combining the equivalences of [47] and [105]. 22 α β γ ǫ δ The directed graph of the NILPs is the basic pattern repeated, with paths moving upwards and to the right, and with weights indicated on the edges. Comparing the weights we get the relations a1 = αβδ b1 = βǫ c1 = δ + ǫγ a2 = 1 b2 = αγ c2 = αβ Combining these we find that a1 a2 + b1 b2 − c1 c2 = 0, so the correspondence only makes sense at ∆ = 0 (and there are really only 4 parameters and not 5 as one might naively assume). 2.4.2. Domino tilings. There is also a prescription to turn six-vertex configurations into domino tilings that is illustrated on Fig. 11(c) [105]. As already mentioned, going from (b) to (c) is nothing but a slightly modified version of the bijection of [47] between NILPs and domino tilings. In order to understand the correspondence of Boltzmann weights, note that patching together the pictures of Fig. 11(c) produces dominoes that span three dotted squares, for example In particular, one half of the domino is contained inside one square. This allows to classify dominoes into four kinds, depending on which half of the square it occupies (these are called north-, west-, south-, and east-going in [47]). Going back to Fig. 11(c), we conclude that a1 , a2 , b1 , b2 can be considered as the Boltzmann weights of the four kinds of dominoes. Furthermore, we have the relations c1 = a1 a2 + b1 b2 c2 = 1 from which we derive as expected a1 a2 + b1 b2 − c1 c2 = 0. Just as plane partitions are dimers on the honeycomb lattice, domino tilings can be considered equivalently as dimers on the square lattice. 2.4.3. Free fermionic five-vertex model. The general five-vertex model is obtained by sending one of the a or b weights to zero while all other weights remain finite; in other words, one simply forbids one of the 6 types of vertices. For a discussion of the general five-vertex model , see for example [83] and in particular its appendix A. In the first part of this section, we choose to send both horizontal and vertical electric fields to minus infinity and a to zero, in such a way that a1 becomes zero. In the representation in terms of north-east going paths, this amounts to forbidding crossings; however, these paths in general interact when they are close to each other. The paths become NILPs (i.e. they only interact through the Pauli principle) only if their weights are products over the edges, which implies that b1 b2 = c1 c2 . This leads us back to the model of section 2.4.1, but with δ sent to zero: what we get this way is the free fermionic five-vertex model, first discussed in [101]. If δ = 0 the NILPs of Fig. 11(b) simply live on a regular square lattice, and of course at this stage we recognize the transfer matrix discussed in section 1.2.4, and illustrated on Fig. 1 (plain 23 Figure 12. From the five-vertex model to dimers or plane partitions. Figure 13. From the five-vertex model to dimers or plane partitions, dual version. lines). In section 1.3 on plane partitions, it was also identified with the transfer matrix of lozenge tilings. To complete the circle of equivalences, we show on Fig. 12 how to go from NILPs to either dimers on the honeycomb lattice or lozenge tilings, following Reshetikhin [93]. There is a second case which is worth mentioning (if only because it will reappear in section 5.2.5): suppose instead that we send b2 to zero. This time the north-east going paths cannot go straight east any more. In this case it is natural to redraw all north-east moves with a right turn as straight lines (not just south-side-goes to east but also west-side-goes-to-north), and this way we recognize the dashed lines of Fig. 1, with a slight modification: the whole picture is distorted in such a way that each path moves one step further to the right (so that north-west becomes north, and north becomes north-east). If we want these paths to be NILPs, we reproduce the weights of 2.4.1 with γ = 0. Finally the correspondence to lozenge tilings/dimers is illustrated on Fig. 13. Note a difference between the models of lozenge tilings corresponding to these two versions of the free fermionic five-vertex models: the vertical spectral parameters flow north-east in the first picture, whereas they flow north-west in the second picture. Ultimately, this is related to two possible inhomogeneous versions of Schur functions (double vs dual [double] Schur functions in the language of [80]). See also the recent work [109] where these lozenge tilings are embedded in a more general square-triangle-rhombus tiling model. 2.5. Domain Wall Boundary Conditions. Domain Wall Boundary Conditions (DWBC) were special boundary conditions which were originally introduced in order to study correlation functions of the six-vertex model [60]. However they are also interesting in their own right. 24 Figure 14. An example of configuration with Domain Wall Boundary Conditions. 2.5.1. Definition. DWBC are defined on a n × n square grid: all the external edges of the grid are fixed according to the rule that vertical ones are outgoing and horizontal ones are incoming. An example is given on Fig. 14. To each horizontal (resp. vertical) line one associates a spectral parameter xi (resp. yj ). The partition function is thus: Zn (x1 , . . . , xn ; y1 , . . . .yn ) = X n Y w(yj /xi ) configurations i,j=1 where w = a, b, c depending on the type of vertex (cf (2.4)). Here we do not allow any electric field for the simple reason that with DWBC (as with any fixed boundary conditions), using the same type of arguments as in the previous section, one can push the effect of the field to the boundary, where it only contributes a constant to the partition function. Remark: the (one-to-many) correspondence of section 2.4.2 sends DWBC six-vertex configurations to domino tilings of the Aztec diamond [46]. 2.5.2. Korepin’s recurrence relations. In [60], a way to compute Zn inductively was proposed. It is based on the following properties: • Z1 = q − q −1 . • Zn (x1 , . . . , xn ; y1 , . . . .yn ) is a symmetric function of the {xi } and of the {yi } (separately). This is a consequence of repeated application of the Yang–Baxter equation (or equivalently of one of the components of the so-called RTT relations): (q yi+1 /yi − q −1 yi /yi+1 )Zn (. . . , yi , yi+1 , . . .) = (q yi+1 /yi − q −1 yi /yi+1 ) yi yi+1 25 x1 y1 Figure 15. Graphical proof of the recursion relation. = = yi yi+1 = ··· = yi yi+1 = (q yi+1 /yi − q −1 yi /yi+1 ) yi yi+1 = (q yi+1 /yi − q −1 yi /yi+1 )Zn (. . . , yi+1 , yi , . . .) yi+1 yi and similarly for the xi . • Zn multiplied by xin−1 (resp. yin−1 ) is a polynomial of degree at most n − 1 in each variable x2i (resp. yi2 ). This is because (i) each variable say xi appears only on row i (ii) a, b are linear combinations of x−1 i , xi and c is a constant and (iii) there is at least one vertex of type c on each row/column. • The Zn obey the following recursion relation: (2.6) Zn (x1 , . . . , xn ; y1 = x1 , . . . , yn ) = (q − q −1 ) n n Y Y (q yj /x1 − q −1 x1 /yj )Zn−1 (x2 , . . . , xn ; y2 , . . . , yn ) (q x1 /xi − q −1 xi /x1 ) i=2 j=2 Since y1 = x1 implies b(y1 /x1 ) = 0, by inspection all configurations with non-zero weights are of the form shown on Fig. 15. This results in the identity. Note that by the symmetry property, Eq. (2.6) fixes Zn at n distinct values of y1 : xi , i = 1, . . . , n. Since Zn is of degree n − 1 in y12 , it is entirely determined by it. 26 2.5.3. Izergin’s formula. Remarkably, there is a closed expression for Zn due to Izergin [40, 39]. It is a determinant formula: (2.7) Qn −1 q − q −1 i,j=1 (xj /yi − yi /xj )(q xj /yi − q yi /xj ) Zn = Q det −1 1≤i<j≤n (xi /xj − xj /xi )(yi /yj − yj /yi ) i,j=1...n (xj /yi − yi /xj )(q xj /yi − q yi /xj ) The hard part lies in finding the formula, but once it is found, it is a simple check to prove that it satisfies all the properties of the previous section. The symmetry under interchange of variables is evident from the structure of the formula, and the recurrence formula follows from looking at the zeroes outside the determinant and the poles inside the determinant: indeed, they must compensate each other for the result to be non-zero at say y1 = x1 , which immediately leads to expanding the determinant on the x1 row and to the recurrence. 2.5.4. Relation to classical integrability and random matrices. The Izergin determinant formula is curious because it involves a simple determinant, which reminds us of free fermionic models. And indeed it turns out that it can be written in terms of free fermions, or equivalently that it provides a solution to a hierarchy of classically integrable PDE, in the present case the two-dimensional Toda lattice hierarchy. We cannot go in any details here but provide a few remarks. Consider a function of two sets of n variables of the form det φ(X, Y ) (2.8) τn (X, Y ) = ∆(X)∆(Y ) where if X = (x1 , . . . , xn ), Y = (y1 , . . . , yn ), then det φ(X, Y ) = det i,j=1,...,n φ(xi , yj ) ∆(X) = Y (xi − xj ) 1≤i<j≤n Certainly Izergin’s formula (2.7) is of this form (up to some prefactors and to xi → x2i , yi → yi2 ); but it is also the case of the Cauchy formula (1.26) (under the form of the Cauchy determinant) and even of Weyl’s formula (1.23) (once divided by ∆(λ); though usually one considers it as a function of the xi only, which results in a tau-function of the KP hierarchy only). It also appears in problems of random matrices, and in particular in the Harish Chandra–Itzykson–Zuber integral [10, 38] (see [106]). τn is of course symmetric by permutation of variables in X, and in Y . We now make use of the following set of bilinear determinant identities: (2.9) n+1 X i=1 ′ ′ det φ(X − xi , Y ) det φ(X + xi , Y ) = X′ m+1 X j=1 det φ(X ′ , Y ′ − yj′ ) det φ(X, Y + yj′ ) ′ where X = (x1 , . . . , xn+1 ), Y = (y1 , . . . , yn ), = (x′1 , . . . , x′m ), Y ′ = (y1′ , . . . , ym+1 ). P P Using as in section 1 the Miwa transformation:3 tq = 1q ni=1 xqi , sq = 1q ni=1 yiq , and similarly for the primed variables, we find after standard contour integration tricks that (2.9) can be rewritten in terms of the τn as I P ′ q du m−n u τm [t − [u], s]τn+1 [t′ + [u], s′ ]e q≥1 (tq −tq )u 2πiu I P ′ q dv n−m v τn [t′ , s′ − [v]]τm+1 [t, s + [v]]e q≥1 (sq −sq )v = 2πiv 3Since we have only a finite number of variables, the correct prescription is to consider a symmetric polynomial in n variables as a linear combination of Schur functions with fewer than n rows. 27 H where [u] = (u−1 , u−2 /2, . . . , u−q /q, . . .), and where the integrals du/(2πiu) have the meaning of picking the constant term of a Laurent series. Expanding this equation in powers of tq − t′q and sq − s′q results in an infinite set of partial differential equations satisfied by the τn . They are the Hirota form of the two-dimensional Toda lattice hierarchy. Inside it, there are two copies of the KP hierarchy corresponding to varying only one set of variables (the tq or the sq ) and keeping n fixed. τn is the tau-function of the hierarchy. In particular, if one expands to first order in t1 − t′1 and set m = n − 1, we find ∂ ∂ ∂ ∂ τn − τn τn (2.10) τn+1 τn−1 = τn ∂t1 ∂s1 ∂t1 ∂s1 which is a form of the Toda lattice equation. There is another representation which is particular useful for the homogeneous limit. Consider the Laplace (or Fourier, we are working at a formal level) transform of φ: ZZ φ(x, y) = dµ(a, b) exa + yb Then one can write τn (X, Y ) = 1 n!∆(X)∆(Y ) Z ··· Z Y n dµ(ai , bi ) i=1 det (exi aj ) i,j=1,...,n det (eyi bj ) i,j=1,...,n This is formally identical to the partition function of a generalized two-matrix model with external fields for both matrices. Next, let us consider the homogeneous limit τn (x, y) of such a function τn (X, Y ) where all xi tend P Pn−1 −1 to x and all yi tend to y. Noting that det(exi aj )/∆(X) ∼ cn ex i ai ∆(a) where cn = ( k=1 k!) (by the usual trick of taking xi = x + iǫ, ǫ → 0) and similarly for the yi , Z Z Y Pn n Pn y x a i j=1 bi i=1 dµ(ai , bi )∆(a)e ∆(b)e τn (x, y) = cn cn+1 · · · i=1 This is a generalized two-matrix model with linear potentials. With an arbitrary potential, the partition function of such a model is known to be a tau-function of the two-dimensional Toda lattice hierarchy. In fact, we have the following fermionic representation, with notations similar to section 1: P Z n xJ + yJ−,1 ⋆ ⋆ dµ(a, b)ψ+ (a)ψ− (b) |0, 0i τn (x, y) ∝ hn, n| e q≥1 +,1 Since we only have here linear potentials i.e. the primary times (the “t1 ”), we shall only recover the first equation of the hierarchy. Let us do so. First note the determinant formula i j ZZ ∂ ∂ i j xa + yb 2 2 φ(x, y) dµ(a, b)a b e = cn det τn (x, y) = cn det i,j=0,...,n−1 ∂xi ∂y j i,j=0,...,n−1 Of course the latter form could have been derived directly from (2.8). Next apply to either of these expressions the Desnanot–Jacobi determinant identity. More precisely, consider the matrix of size n + 1 and write that its determinant times the determinant of the sub-matrix of size n − 1 with last two rows and columns removed equals the difference of the two possible products of determinants of sub-matrices of size n with one row, one column, removed and the other row, other column removed (among the last two rows and columns). The result is: ∂ ∂ ∂ ∂ (2.11) n2 τn+1 τn−1 = τn τn − τn τn ∂x ∂y ∂x ∂y (the factor of n2 takes care of the c2n ). 28 Finally, in the special case that φ(x, y) only depends on x−y, then the previous formulae simplify. The measure dµ(a, b) is concentrated on a = −b and we have Z Z Y n 1 τn (X, Y ) = dµ(ai ) det (exi aj ) det (e−yi aj ) ··· i,j=1,...,n i,j=1,...,n ∆(X)∆(Y ) i=1 which is a generalized one-matrix model with external field. In the homogeneous limit, τn (x, y) becomes a function of a single variable t = x − y and we can write Z Z Y n Pn dµ(ai )∆(a)2 et i=1 ai (2.12) τn (t) = · · · i=1 which is a one-matrix model with linear potential. With an arbitrary potential, the partition function of such a model is known to be a tau-function of the Toda chain hierarchy. Writing as before i+j ZZ 1 ∂ 1 i+j ta τn (t) = det det φ(t) dµ(a)a e = (n!)2 i,j=0,...,n−1 (n!)2 i,j=0,...,n−1 ∂ti+j and applying the Jacobi–Desnanot identity we obtain (2.13) n2 τn+1 τn−1 = τn τn′′ − τn′ 2 which is a form of the Toda chain equation. 2.5.5. Thermodynamic limit. In [61], the Toda chain equation (2.13) above was used to derive the asymptotic behavior of the partition function of the six-vertex model with DWBC in two of its three phases: ferroelectric and disordered. These are the two phases where we expect the limit n → ∞ to be smooth, as we shall explain. In this case, making the Ansatz that the free energy is 2 extensive, we plug the asymptotic expansion τn ∼ e−n f into the Toda chain equation and are left with a simple differential equation to solve: (2.14) f ′′ = e2f In fact, this is the simplest reasonable Ansatz that is compatible with (2.13), so that any solution of (2.13) with a smooth large n limit will be governed by the differential equation (2.14). This is how the computation of the bulk free energy of the six-vertex model with DWBC for ∆ > −1 is performed in [61]. It is much simpler than the computation for PBC and the result is given in terms of elementary functions, and is thus different from that of PBC. Explicitly, we find ( max(a, b) ∆>1 2 (2.15) Zn1/n → π/γ i a b cos πt/γ |∆| < 1, q = −e−iγ (γ > 0), x = ei(t+γ) Note the obvious interpretation of the result in the ferroelectric regime – a frozen phase where almost all arrows are aligned. In the anti-ferroelectric phase ∆ < −1 one expects a less smooth large n limit because the alternation of arrows that is favored in this phase will interact with the boundaries. This statement is made more precise in [105], where matrix model techniques are used to compute the large n limit of (2.12) (which essentially boil down to an appropriate saddle point analysis of it). And indeed one finds that log τn has an oscillating term of order 1, which explains that the simple differential equation (2.14) cannot account for the asymptotic behavior. In any case, in all phases one finds a result that is different from that of PBC. The explanation of this phenomenon is that the six-vertex model suffers from a strong dependence on boundary conditions due to the constraints imposed by arrow conservation. In particular there is no thermodynamic limit in the usual sense (i.e. independently of boundary conditions). In [104] it was 29 0 0 0 0 1 −1 Figure 16. From six-vertex to ASMs. suggested more precisely that the six-vertex model undergoes spatial phase separation, similarly to plane partitions [13] and other dimer models [53]. In other words, even far from the boundary of the system, the system loses any translational-invariance and the physical behavior around a given point is a function of the local polarization: as such, the model can have several (possibly, all) of the three phases coexisting in different regions. This was motivated by some numerical evidence, as well as by the exact result at the free fermion point ∆ = 0, at which the arctic circle theorem [46] applies: the boundary between ferroelectric and disordered phases is known exactly to be an ellipse (a circle for a = b) tangent to the four sides of the square. So the apparent simplicity of the computation of the bulk free energy for DWBC conceals a complicated physical picture. Since then, there has been a considerable amount of work in this area. There has been more numerical work [1]. Some of the results of [61] have been proven rigorously and extended using sophisticated machinery in the series of papers [6, 8, 7] by Bleher et al. Finally, the curve separating phases has been studied in the work of Colomo and Pronko [14, 15], and recently they proposed equations for this curve in the cases a = b, ∆ = ±1/2 [16]. The point ∆ = 1/2 is of special interest: it corresponds to all weights equal, and is the original ice model. It is also the subject of the next section. 2.5.6. Application: Alternating Sign Matrices. Alternating Sign Matrices are an important class of objects in modern combinatorics [77]. They are defined as follows. An Alternating Sign Matrix (ASM) is a square matrix made of 0s, 1s and -1s such that if one ignores 0s, 1s and -1s alternate on each row and column starting and ending with 1s. For example, 0 0 1 0 1 0 0 0 1 0 0 −1 1 0 1 0 is an ASM of size 4. The enumeration of ASMs is a famous problem with a long history, see [9]. Here we simply note that ASMs are in fact in bijection with six-vertex model configurations with DWBC [62]. The correspondence is quite simple and is summarized on Fig. 16. For example, Fig. 14 becomes the 4 × 4 ASM above. We can therefore reinterpret the partition function of the six-vertex model with DWBC as a weighted enumeration of ASMs. It it natural to set the weight of all zeroes to be equal (a = b), which leaves us with only one parameter c/a, the weight of a ±1. In fact here we shall consider only the pure enumeration problem that is all weights equal. We thus compute ∆ = 1/2 and q = eiπ/3 , and then xi = q, yj = 1 so that the three weights are w(xi /yj ) = q − q −1 . At this stage there are several options. Either one tries to evaluate directly the formula (2.7); since the determinant vanishes in the homogeneous limit where all the xi or yj coincide, this is a somewhat involved computation and is the content of Kuperberg’s paper [62]. 30 There is however a much easier way, discovered independently by Stroganov [98] and Okada [85]. It consists in identifying Zn at q = eiπ/3 with a Schur function. Consider the partition λ(n) = (n − 1, n − 1, n − 2, n − 2, . . . , 1, 1), that is the Young diagram (2.16) λ (n) = n−1 z }|. . . ... .. ... . . . . .. . { sλ(n) (z1 , . . . , z2n ) is a polynomial of degree at most n − 1 in each zi (use (1.24)) and, satisfies the following (2.17) sλ(n) (z1 , . . . , zj = q −2 zi , . . . , zn ) = 2n Y k=1 k6=i,j (zk − q 2 zi )sλ(n−1) (z1 , . . . , ẑi , . . . , ẑj , . . . , z2n ) where the hat means that these variables are skipped (start from (1.13), find all the zeroes as zj = q 2 zi and then set zi = zj = 0 to find what is left). This looks similar to recursion relations (2.6). After appropriate identification one finds: n Y Zn (x1 , . . . , xn ; y1 , . . . , yn )q=eiπ/3 = (−1)n(n−1)/2 (q − q −1 )n (q xi yi )−(n−1) i=1 sλ(n) (q 2 x21 , . . . , q 2 x2n , y12 , . . . , yn2 ) Note that Zn possesses at the point q = eiπ/3 an enhanced symmetry in the whole set of variables {q x1 , . . . , q xn , y1 , . . . , yn }. Finally, setting xi = q −1 and yj = 1 and remembering that this will 2 give a weight of (q − q −1 )n to each ASM, one concludes that the number of ASMs is given by −n(n−1)/2 An = 3 −n(n−1)/2 sλ(n) (1, . . . , 1) = 3 | {z } Y (n) λi 1≤i<j≤2n 2n (n) − i − λj +j j−i Simplifying the product results in (2.18) An = n−1 Y i=0 (3i + 1)! = 1, 2, 7, 42, 429 . . . (n + i)! which is a sequence of numbers we have encountered before! In fact, the first proof of formula (2.18), due to Zeilberger [102], amounts to showing (non-bijectively) that the number of ASMs is the same as the number of TSSCPPs. These are the ASMs of size 1, 2, 3 (+ = 1, − = −1): + + 0 0 + + 0 0 0 + 0 0 0 + + 0 0 0 0 + 0 + 0 0 + 0 + 0 0 0 0 + 0 + + 0 0 + 0 0 0 + + 0 0 31 0 + 0 +−+ 0 + 0 0 0 + + 0 0 0 + 0 0 0 + 0 + 0 + 0 0 (a) (b) Figure 17. Vertices of (a) the CPL model and (b) the FPL model. As a check, one can take n → ∞ in (2.18), and using Stirling’s formula one finds √ 3 3 1/n2 An → 4 1/n2 This is to be compared with for γ = 2π/3, where we find Zn √ (2.15) 2 n agree considering Zn = ( 3i/2) An . → 9i/8. The two formulae 3. Loop models and Razumov–Stroganov conjecture 3.1. Definition of loop models. Loop models are an important class of two-dimensional statistical lattice models. They display a broad range of critical range of critical phenomena, and in fact many classical models are equivalent to a loop model. The critical exponents, formulated in the language of loop models, often acquire a simple geometric meaning; and many methods have been used to study their continuum limit, including the Coulomb Gaz approach [82], Conformal Field Theory [5] and more recently the Stochastic Löwner Evolution [65, 66, 67]. Here we are of course more interested in their properties on a finite lattice (in relation to combinatorics) and in the use of integrable methods. We shall introduce two classes of loop models on the square lattice, which turn out to be both closely related to the six-vertex model. Then we shall discuss a very non-trivial connection between these two loop models (the Razumov–Stroganov conjecture). Let us first discuss common features of the two models. Their configurations consist of loops living on the edges of the square lattice. The most important feature is the non-local Boltzmann weight produced by assigning a fugacity of τ to closed loops. Here τ is a real parameter (usually called n, due to the connection with the O(n) model – however for various reasons we avoid this notation here). This can be supplemented by possible local weights for the various configurations around a given vertex, see Fig. 17. 3.1.1. Completely Packed Loops. Configurations of Completely Packed Loops (CPL) consist of nonintersecting loops occupying every edge of the square lattice, which produces two possibilities at each vertex, represented on Fig. 17(a). Besides the weight τ of each closed loop, one can introduce a local weight of u for one of the two types of CPL vertices, say NE/SW loops. The model is known to be critical for |τ | < 2, and its continuum limit is described by a theory γ2 , where τ = 2 cos γ, 0 < γ < π. with central charge c = 1 − 6 π(π−γ) 3.1.2. Fully Packed Loops: FPL and FPL2 models. Configurations of Fully Packed Loops (FPL) consist of non-intersecting loops such that there is exactly one loop at each vertex, which results in the six possibilities described on Fig. 17(b). One can then give a weight of τ to each closed loop (FPL model). However, one can do better: noting that the empty edges also form loops (dashed lines on the figure), one can put them on the same footing as occupied edges and assign them a fugacity too, say τ̃ . This more general model is usually called FPL2 model. 32 odd even Figure 18. From six-vertex to FPLs. One reason that the FPL2 is interesting is the following: the FPL model is not integrable for τ 6= 1 (the very special case τ = 1 is of interest to us and will be considered below). However, the FPL2 is integrable for τ = τ̃ , that is if the two types of loops are given equal weights. This was shown using Coordinate Bethe Ansatz in [21] and then rederived using Algebraic Bethe Ansatz in [43]. For generic values of τ , τ̃ , the Coulomb Gaz approach provides non-rigorous arguments to identify γ2 − the continuum field theory, see [41], and allows to compute the central charge to be c = 3−6 π(π−γ) 2 γ̃ 6 π(π−γ̃) , where τ = 2 cos γ and τ̃ = 2 cos γ̃, 0 < γ, γ̃ < π. In particular the FPL model has central 2 γ , which is one more than the corresponding CPL model. charge c = 2 − 6 π(π−γ) We shall not discuss the possibility of adding local weights in detail. Let us simply note that even if we impose rotational invariance of local Boltzmann weights, we can introduce an energy cost for 90 degrees turns of the loops, which amounts to giving them a certain amount of bending rigidity. Such a model was studied numerically in [42]. 3.2. Equivalence to the six-vertex model and Temperley–Lieb algebra. 3.2.1. From FPL to six-vertex. The relation between six-vertex model and FPL model is rather limited, so we treat it first. The limitation comes from the fact that one cannot assign an actual weight to the loops, so that we obtain a τ = 1 model (with only local weights). The correspondence between configurations is one-to-one: starting from the six-vertex model side, one imposes that at every vertex, arrows pointing in the same direction should be in the same state (occupied or empty) on the FPL side. This forces us to distinguish odd and even sub-lattices, and leads to the rules of Fig. 18. For rotational invariance of the FPL weights one should have a = b. c/a then plays the role of rigidity parameter of the loops mentioned in section 3.1.2. The rest of this section is devoted to the equivalence of CPL and six-vertex models. 3.2.2. From CPL to six-vertex. An example is shown on Fig. 19. Start from a CPL configuration. The (unoriented) loops carry a weight of τ . A convenient way to make the latter weight local is to turn unoriented loops into oriented loops: each configuration is now expanded into 2# loops configurations with every possible orientation of the loops. The weight of a 90 degrees turn is chosen to be ω ±1/4 , where τ = ω + ω −1 . 33 → + ··· → + ··· Figure 19. From CPLs to six-vertex. Finally we forget about the original loops, retaining only the arrows. We note that the arrow conservation is automatically satisfied around each vertex: we thus obtain one of the six vertex configurations. a= = =u b= = =1 c1 = + = u ω 1/2 + ω −1/2 c2 = + = ω 1/2 + u ω −1/2 Note that if u = 1 all weights become rotationally invariant and a = b, c1 = c2 . Finally, one checks that the formula ∆ = −τ /2 holds (equivalently q = −ω), u playing the role of spectral parameter. In particular the critical phase |∆| < 1 corresponds to |τ | < 2. Remark: this construction only works in the plane. On the cylinder or on the torus we have a problem: there are non-contractible loops which according to the prescription above get a weight of 2. This issue will reappear in the section 3.2.4 under the form of the twist. It explains the discrepancy of central charges between 6-vertex model (c = 1) and CPL (c < 1). 3.2.3. Link Patterns. In order to understand this equivalence at the level of transfer matrices, one needs to introduce an appropriate space of states for the CPL model. We now assume for simplicity that L is even, L = 2n. Define a link pattern of size 2n to be a non-crossing pairing in a disk of 2n points lying on the boundary of the disk. Strictly equivalently we can map the disk to the upper half-plane and “flatten” link patterns to pairings inside the upper half-plane of points on its boundary (a line). We shall switch from one description to the other depending on what is more convenient. The points are labelled from 1 to 2n; in the half-plane, they are always ordered from left to right, whereas in the disk the location of 1 must be chosen, after which the labels increase counterclockwise. Denote the set of link patterns of size 2n by P2n . The number of such link patterns is cn = the so-called Catalan number. 34 (2n)! n!(n+1)! , Example: in size L = 6, there are 5 link patterns: 5 5 5 5 5 6 4 6 4 6 4 6 4 6 4 1 3 1 3 1 3 1 3 1 3 2 2 2 2 2 3.2.4. Periodic Boundary Conditions and twist. Suppose that we consider the CPL model with periodic boundary conditions in the horizontal direction, with a width of L = 2n. We can define a transfer matrix with indices living in the set of link patterns P2n as follows. Consider appending a row of the CPL model to a link pattern; this way one produces a new link pattern: 7 6 7 8 5 −→ 1 4 6 8 5 1 4 2 2 3 3 The transfer matrix Tππ′ is then the sum of weights of CPL rows such that the pattern π ′ is turned into the pattern π. The weights are calculated as follows: first one takes the product over each plaquette of the local weights; and then one multiplies by τ to the power the number of loops that we have created. What is the precise correspondence between the space of link patterns CP2n and the space of L spins C2 which relates the transfer matrix of the six-vertex model and the newly defined one for the CPL model? We start from the equivalence described in the section 3.2.2. The basic idea is to orient the loops. So we start from a link pattern and add arrows to each “loop” (pairing of points). Forgetting about the original link pattern we obtain a collection of 2n up or down arrows, which form a state of the 6-vertex model in the transfer matrix formalism. To assign weights it is convenient to think of the points as being on a straight line with the loops emerging perpendicularly: this way each loop can only acquire a weight of ω ±1/2 , depending on whether it is moving to the right of to the left. For example, in size L = 2n = 4, 1 1 2 2 3 3 4 4 =ω 1 2 3 4 + 1 2 3 4 + 1 2 3 4 +ω −1 −1 = ω ↑1 ↑2 ↓3 ↓4 + ↑1 ↓2 ↑3 ↓4 + ↓1 ↑2 ↓3 ↑4 +ω =ω + + +ω −1 1 2 3 4 = ω ↑1 ↓2 ↑3 ↓4 1 2 3 4 + ↑1 ↓2 ↓3 ↑4 1 2 3 4 + ↓1 ↑2 ↑3 ↓4 +ω −1 1 2 3 4 ↓1 ↓2 ↑3 ↑4 1 2 3 4 ↓1 ↑2 ↓3 ↑4 There is only one problem with this correspondence: it is not obviously compatible with periodic boundary conditions. We would like to identify a loop from i to j, i < j and a loop from j to i + L, j < i + L. This is only possible if we assume that ↑i+L = ω ↑i , ↓i+L = ω −1 ↓i , i.e. we impose twisted 35 z boundary conditions on the six-vertex model. In the notations of (2.3) the twist is Ω = ω σ : it corresponds to an imaginary electric field. This mapping from the space of link patterns (of dimension cn ) to that of sequences of arrows (of dimension 22n ) is injective; so that the space of link patterns is isomorphic to a certain subspace 2n C2 . The claim, which we shall not prove in detail here but which is a natural consequence of the general formalism is that the transfer matrix (2.3) of the six-vertex model with the twist defined above leaves invariant this subspace and, once restricted to it, is identical to the transfer matrix of our loop model up to this isomorphism, the correspondence of weights being the same as in section 3.2.2 (in particular, ∆ = −τ /2). The connection between CPL and six-vertex models is deep, in the sense that they are based on the same algebraic structure, the Temperley–Lieb algebra and the associated solution of the Yang–Baxter equation, but in different representations. This is what we briefly discuss now. These definitions will serve again when we study the quantum Knizhnik–Zamolodchikov equation (section 4). 3.2.5. Temperley–Lieb and Hecke algebras. The Temperley–Lieb algebra of size L and with parameter τ is given by generators ei , i = 1, . . . , L − 1, and relations: e2i = τ ei (3.1) ei ei±1 ei = ei ei ej = ej ei |i − j| > 1 It is a quotient of the Hecke algebra, i.e. the ei satisfy the less restrictive relations (3.2) e2i = τ ei ei ei+1 ei − ei = ei+1 ei ei+1 − ei+1 ei ej = ej ei |i − j| > 1 Note that in Hecke (not Temperley–Lieb!), there is a symmetry ei ↔ τ − ei . The Hecke algebra is itself a quotient of the braid group algebra: if τ = −(q + q −1 ) (q is thus a free parameter), then the ti = q −1/2 ei + q 1/2 satisfy the relations ti ti+1 ti = ti+1 ti ti+1 ti tj = tj ti |i − j| > 1 We are interested in two representations of the Temperley–Lieb algebra. The first one is the representation on the “space of spins”, that is the same space (C2 )⊗L of sequences of up/down arrows on which the six-vertex transfer matrix acts. It is given by making ei act on the ith and (i + 1)st copies of C2 in the tensor product, with matrix 0 0 0 0 0 −q 1 0 (3.3) ei = −1 0 1 −q 0 0 0 0 0 It may be expressed in terms of Pauli matrices as (3.4) ei = 1 x x y z z σi σi+1 + σiy σi+1 + ∆(σiz σi+1 − 1) + h(σi+1 − σiz ) 2 where h = 12 (q − q −1 ) and the σi are the Pauli matrices at site i, The second representation of Temperley–Lieb can be defined purely graphically. We now assume as before that L is even, L = 2n. In order to define the action of Temperley–Lieb generators ei on the space of link patterns CP2n (vector space with canonical basis the |πi indexed by link patterns), ; then the relations of the Temperley–Lieb it is simpler to view them graphically as ei = i i+1 36 algebra, as well as the representation on the space of link patterns, become natural on the picture; for example, we find e1 = 1 2 3 4 5 = 6 1 1 e2 2 3 4 5 = 1 2 3 4 5 2 3 4 5 6 6 =τ 6 1 1 2 3 4 5 2 3 4 5 6 6 As before, the role of the parameter τ is that each time a closed loop is formed, it can be erased at the price of a multiplication by τ . The two representations we have just defined are of course related: the transformation of section 3.2.4 makes the representation on link patterns a sub-representation of the one on spins. Finally, we need to define affine versions of Temperley–Lieb and Hecke algebras. It is convenient to do so by starting from their non-affine counterparts and adding an extra generator ρ, as well as relations ρei = ei+1 ρ, i = 1, . . . , L − 1, and ρL = 1. Note that this allows to define a new element eL = ρeL−1 ρ−1 = ρ−1 e1 ρ such that all defining relations of the algebra become true modulo L. In fact, a more standard approach would be to introduce only eL and not ρ itself. Adding ρ leads to a slightly extended affine Hecke/Temperley–Lieb algebra, which is more convenient for our purposes (see the discussion in [87]). In the link pattern representation, ρ simply rotates link patterns counterclockwise: 7 6 8 7 5 ρ 6 8 5 1 4 = 1 4 2 3 2 3 In the spin representation, it rotates the factors of the tensor product forward one step and twists the last one (that moves to the first position) by Ω. Once again, for these two representations to z be equivalent, one needs Ω to be of the form Ω = ω σ where ω = −q. As an application, consider the Hamiltonian HL , obtained as the logarithmic derivative of the transfer matrix TL (x) of (2.5) evaluated at x = 1. We find that with periodic boundary conditions, it is simply given up to additive and multiplicative constants by HL = − L X ei i=1 In the spin representation, using (3.4), we recognize in HL the Hamiltonian of the XXZ spin chain (with the so-called ferromagnetic sign convention). So the Hamiltonian of the loop model, which has the same form, is equivalent to the XXZ spin chain Hamiltonian, but with twisted periodic ± = ω ±2 σ1± . boundary conditions, which in terms of Pauli matrices means that σL+1 37 1 2 3 4 ... ... 2n Figure 20. The CPL model on a cylinder. 3.3. Some boundary observables for loop models. Here we consider the CPL model at τ = 1 with some specific boundary conditions which will play an important role since the observables we shall compute live at the boundary. Several geometries are possible and lead to interesting combinatorial results [17], but here we only consider the case of a cylinder. 3.3.1. Loop model on the cylinder. We consider the model of Completely Packed Loops (CPL) on a semi-infinite cylinder with a finite even number of sites L = 2n around the cylinder, see Fig. 20. It it convenient to draw the dual square lattice of that of the vertices, so that the cylinder is divided and . into plaquettes. Each plaquette can contain one of the two drawings We furthermore set τ = 1 (or q = e2πi/3 ), that is we do not put any weights on the loops. There are no more non-local weights, and in fact plaquettes are independent from each other. So we can reformulate this model as a purely probabilistic model, in which one draws independently at and 1 − p for . random each plaquette, with say probability p for Finally, we define the observables we are interested in. We consider the connectivity of the boundary points, i.e. the endpoints of loops (which are in this case not loops but paths) lying on the the bottom circle. We encode them into link patterns (see section 3.2.3). In the present context, they can be visualized as follows. Project the cylinder onto a disk in such a way that the boundaries coincide and the infinity is somewhere inside the disk. Remove all loops except the boundary paths. Up to deformation of these resulting paths, what one obtains is a link pattern. The probabilities of occurrence of the various link patterns can be encoded as one vector with cn entries: X Ψπ |πi ΨL = π∈P2n where P2n is the set of link patterns of size 2n and Ψπ is the probability of link pattern π. 3.3.2. Markov process on link patterns. We now show that ΨL can be reinterpreted as the steady state of a Markov process on link patterns. This is easily understood by considering a transfer matrix formulation of the model. As in section 3.2.4, let us introduce the transfer matrix: it corresponds here to creating one extra row to the semi-infinite cylinder, and encoding not the actual plaquettes but the effect of the new plaquettes on the connectivity of the endpoints. That is, Tπ,π′ (p) is the probability that starting from a configuration of the cylinder whose endpoints are 38 connected via the link pattern π ′ and adding a row of plaquettes, one obtains a new configuration whose endpoints are connected via the link pattern π. This form a cn × cn matrix TL (p). This transfer matrix is actually stochastic in the sense that X (3.5) ∀π ′ Tπ,π′ (p) = 1 π∈P2n which expresses the conservation of probability. This is of course a special feature of the transfer matrix at τ = 1. Note that (3.5) says that TL (p)T has eigenvector (1, . . . , 1) with eigenvalue 1. The matrix TL (p) has non-negative entries; it is easy to show that it is primitive (the entries of TL (p)n are positive). These are the hypotheses of the Perron–Frobenius theorem. Therefore, TL (p) possesses a unique eigenvector ΨL with positive entries; the corresponding eigenvalue is positive and is larger in modulus than all other eigenvalues. Now the theorem also applies to TL (p)T and by uniqueness we conclude that the largest eigenvalue of TL (p) and of TL (p)T is 1. In conclusion, we find that the eigenvector with positive entries of TL (p), which with a bit of foresight we call ΨL again, satisfies (3.6) TL (p)ΨL = ΨL (In fact the whole reasoning in the previous paragraph is completely general and applies to any Markov process, ΨL being up to normalization the steady state of the Markov process defined by TL (p).) Two more observations are needed. Firstly, (3.6) is clearly satisfied by the vector of probabilities that we defined in the previous paragraph (the semi-infinite cylinder being invariant by addition of one extra row); it is in fact defined uniquely up to normalization by (3.6). This explains that we have used the same notation. Secondly, ΨL is in fact independent of p. The easiest way to see this is to note that p now plays the role of spectral parameter (explicitly, with the conventions of the six-vertex R-matrix, −1 x−1 ). In particular we conclude that, as in the six-vertex model, [TL (p), TL (p′ )] = 0, so p = qq x−q x−1 −q −1 x that these transfer matrices have a common Perron–Frobenius eigenvector. 3.3.3. Properties of the steady state: some empirical observations. We begin with an example in size L = 2n = 8. By brute force diagonalization of the stochastic matrix TL one obtains the vector ΨL of probabilities: 7 1 Ψ8 = 42 3 + 42 6 7 8 5 1 4 6 7 8 5 1 4 + 6 7 8 5 1 4 + 6 8 5 1 4 + 2 3 2 3 2 3 2 3 7 6 7 6 7 6 7 6 8 5 8 5 + 1 4 2 3 8 5 + 1 4 2 3 1 4 2 39 8 5 1 4 + 3 2 3 7 6 8 7 5 + 7 + 42 7 8 5 + 1 6 4 1 4 3 2 3 7 6 7 6 5 1 4 8 5 1 4 + 2 3 8 7 5 + 2 8 6 2 3 8 5 1 4 + 1 6 4 2 3 2 3 We recognize some of our favorite numbers An , namely 7 and 42. In fact, Batchelor, de Gier and Nienhuis [3] conjectured the following properties for all system sizes L = 2n: (1) The smallest probability is 1/An , and corresponds to link patterns with all parallel pairings. (2) All probabilities are integer multiples of the smallest probability. (3) The largest probability is An−1 /An , and correspond to link patterns which pair nearest neighbors. By now all these properties have been proven [23, 28], as will be discussed in section 4. 3.3.4. The general conjecture. A question however remains: according to property 2 above, if one multiplies the probabilities by An , we obtain a collection of integers. The smallest one is 1 and the largest one is An−1 , but what can we say about the other ones? Recall that An also counts the number of six-vertex model configurations with DWBC. Furthermore, we showed that there is a one-to-one correspondence between six-vertex model configurations and FPL configurations (cf section 3.2.1). In this correspondence, the DWBC become alternating occupied and empty external edges for FPL configurations (in short, FPLs). Let us now draw explicitly the 42 FPLs of size 4 × 4: 40 Interestingly, we find that the reformulation in terms of FPLs allows to introduce once again a notion of connectivity. Indeed, there are 2n occupied edges on the exterior square and they are paired by the FPL. We can therefore count separately FPLs with a given link pattern π; let us denote this by Aπ . The Razumov–Stroganov conjecture [90] then states that Ψπ = Aπ An thus relating two different models of loops (CPL and FPL) with completely different boundary conditions. And even though both models are equivalent to the six-vertex model, the values of ∆ are also different (they differ by a sign). The Razumov–Stroganov conjecture remains open, though some special cases have been proved, see 4.3.2 below. The relation to the conjectured properties of the previous section is as follows. It is easy to show that if π is a link pattern with all pairings parallel, then there exists a unique FPL configuration with connectivity π. Thus the RS conjecture implies property (1). Furthermore, since all Aπ are integer, it obviously implies property (2). Property (3) however remains non-trivial, since even assuming the RS conjecture it amounts to saying that Aπ = An−1 in the case of the two link patterns π that pair nearest neighbors, which has not been proven. 4. The quantum Knizhnik–Zamolodchikov equation We now introduce a new equation whose solution will roughly correspond to a double generalization of the ground state eigenvector ΨL of the loop model introduced above: (i) it contains inhomogeneities and (ii) it is a continuation of the original eigenvector to an arbitrary value of q, the original value being q = e2πi/3 . 4.1. Basics. 4.1.1. The qKZ system. Let L = 2n be an even positive integer. Introduce once again the R-matrix, but this time rotated 45 degrees and which acts a bit differently than before. Namely, it acts on the vector space CP2n spanned by link patterns, in the following way: (4.1) Ři (z) = φ(z) (q −1 − qz) + (1 − z)ei q −1 z − q i = 1, . . . , L − 1 where we refer to section 3.2.5 for the definition of ei . φ(z) is a scalar function that we do not need to specify explicitly here, see section 4.1.2 for a discussion. Redrawing slightly the latter as ei = 41 , and similarly the identity as 1 = , we recognize the two (rotated) CPL plaquettes. Note that in this section, it is convenient to use spectral parameters z that are the squares of our old spectral parameters x. Indeed, using the equivalence to the six-vertex model described in section 3.2.4, which amounts to the representation (3.3) for the Temperley–Lieb generators (acting on the ith and (i + 1)st spaces), we essentially recover the R-matrix of the six-vertex model, cf (2.1,2.4), after the change of variables z = x2 : q x − q −1 x−1 0 0 0 φ(x2 ) 0 (q − q −1 )x−1 x − x−1 0 Ř(x) = −1 −1 −1 −1 0 x−x (q − q )x 0 qx −q x −1 −1 0 0 0 qx−q x provided one performs the following transformation: Ř(z) ∝ Pxκ/2 R(x)x−κ/2 where P permutes the factors of the tensor product, and κ = diag(0, 1, −1, 0). The Yang–Baxter equation can be rewritten as follows: Ři (z)Ři+1 (z w)Ři (w) = Ři+1 (w)Ři (z w)Ři (z) i = 1, . . . , L − 1 We also require that φ(z)φ(1/z) = 1, so that Ři satisfies the unitary equation: Ři (z)Ři (1/z) = 1 Consider now the following system of equations for ΨL , a vector-valued function of the z1 , . . . , zL , q, q −1 : (i = 1, . . . , L − 1) (4.2) (4.3) Ři (zi+1 /zi )ΨL (z1 , . . . , zL ) = ΨL (z1 , . . . , zi+1 , zi , . . . , zL ) ρ−1 ΨL (z1 , . . . , zL ) = κ ΨL (z2 , . . . , zL , s z1 ) where it is recalled that ρ rotates link patterns counterclockwise,4 and κ is a constant that is needed for homogeneity. s is a parameter of the equation: if one sets s = q 2(k+ℓ) with k = 2 (technically, d this is the dual Coxeter number of the underlying quantum group Uq (sl(2))), then ℓ is called the level of the qKZ equation. (4.2) is the exchange equation, which is ubiquitous in integrable models. (4.3) has something to do with our (periodic) boundary conditions, or equivalently with an appropriate affinization of the underlying algebra. This system of equations first appeared in [96] in the study of form factors in integrable models. It is not what is usually called the quantum Knizhnik–Zamolodchikov (qKZ) equation; the latter was introduced in the seminal paper [35] as a q-deformation of the Knizhnik–Zamolodchikov (KZ) equation (qKZ is to quantum affine algebras what KZ is to affine algebras). The usual qKZ equation involves the operators Si , which Si can be defined pictorially as ··· Si = · · · i where the empty box is just the “face” graphical representation of the R-matrix (dual to the “vertex” representation we have used for the six-vertex model): q −1 − q z 1−z + −1 −1 q z−q q z−q and the spectral parameters z to be used in Si are as follows: for the box numbered j, zj /zi if j > i or zj /(s zi ) if j < i. Loosely, Si is the “scattering matrix for the ith particle”. = 4We use ρ−1 in the l.h.s. because (ρ−1 Ψ ) = Ψ L π ρ(π) ! 42 Alternatively, Si can be expressed as a product of Ři and ρ: (4.4) Si = Ři−1 (zi−1 /(s zi )) · · · Ř2 (z2 /(s zi )Ř1 (z1 /(s zi ))ρŘL−1 (zL /zi ) · · · Ři+1 (zi+2 /zi )Ři (zi+1 /zi ) The quantum Knizhnik–Zamolodchikov equation can then be written (4.5) Si (z1 , . . . , zL )ΨL (z1 , . . . , zi , . . . , zL ) = ΨL (z1 , . . . , s zi , . . . , zL ) It is a simple exercise to check using (4.4) that the system (4.2–4.3) implies the qKZ equation (4.5). Naively, the converse is untrue. However one can show that, up to some linear recombinations, a complete set of (meromorphic) solutions of (4.5) can always be reduced to a complete set of solutions of (4.2–4.3), see [11]. 4.1.2. Normalization of the R-matrix. In the usual setting of the qKZ equation, the normalization φ(z) of the R-matrix is chosen such that it satisfies an additional constraint, the crossing symmetry, as well as certain analyticity requirements. There is however a certain freedom in choosing this normalization. Indeed, consider a solution Ψ of the system (4.2,4.3) and redefine Y Ψ̃ = f (zj /zi ) Ψ 1≤i<j≤L where f (z) is some scalar function. Then Ψ̃ satisfies the same system of equations: the first equation remains formally identical, but with the normalization of the R-matrix modified from φ(z) to φ̃(z) = φ(z)f (z)/f (1/z). For the second equation to remain true imposes a constraint on the function f (z), so that we find f (z)/f (1/z) = φ̃(z)/φ(z) f (s z) = f (1/z) With reasonable analyticity conditions on φ̃/φ, this system of equations can be solved (e.g. if s < 1, Q∞ Q i −1 i writing φ̃(z)/φ(z) = f0 (z)/f0 (1/z), one finds f (z) = ( ∞ i=1 f0 (s z)) ). i=0 f0 (s /z) In what follows we shall only be interested in homogeneous polynomial solutions of the system (4.2–4.3). In this case the most convenient normalization is to set φ = 1, which we shall adopt from now on. Note that iterating (4.3) L times implies that κL sd = 1 where d is the degree of Ψ. 4.1.3. Relation to affine Hecke algebra. The role of the affine Hecke algebra (and even double affine Hecke algebra) was emphasized in the work of Cherednik [12] in relation to Macdonald polynomials and the quantum Knizhnik–Zamolodchikov equation, and in the work of Pasquier [87] on integrable models. In particular, its use in the context of the Razumov–Stroganov conjecture was advocated by Pasquier [89]. Start from the qKZ system (4.2–4.3), and rewrite it in such a way that the action on the finitedimensional part (on the space of link patterns) is separated from the action on the variables (on the space of polynomials of L variables). Start with (4.2); after simple rearrangements it becomes (4.6) (τ − ei )ΨL = −(q zi − q −1 zi+1 )∂i ΨL where ∂i ≡ zi+11−zi (si − 1) and si is the operator that switches variables zi and zi+1 , so that the l.h.s. only acts on the polynomial part of ΨL , whereas the r.h.s. only acts on link patterns. The operators τ − êi := −(q zi − q −1 zi+1 )∂i acting on the space of polynomials form a representation of the Hecke algebra (with parameter τ ); i.e. they satisfy the relations of (3.2). Equivalently, the êi satisfy them and form a second set of generators. 43 Figure 21. From a Young diagram to a link pattern: in this example, from the partition (2, 1, 1) to the pairings (1, 2), (3, 8), (4, 5), (6, 7). As to (4.3), it is already in a separated form; and one can add an extra operator on the space of polynomials, the one that appears in the r.h.s. of (4.3), that is the cyclic shift ρ̂ of spectral parameters z1 7→ z2 7→ · · · 7→ zL 7→ s z1 (with the multiplication by κ included). The êi together with ρ̂ generate a representation of the affine Hecke algebra. We can write formally: (4.7) ei ΨL = êi ΨL (4.8) ρ−1 ΨL = ρ̂ΨL We can now interpret (4.7,4.8) as follows: we have on the one hand a representation of the affine Hecke algebra on CP2n , the space of link patterns (with generators ei and ρ); and on the other hand a representation of the same algebra on C[z1 , . . . , zL ], the polynomials of L variables (the êi and ρ̂). ΨL provides a bridge between these two representations: it is essentially an invariant object in the tensor product of the two, that is it provides a sub-representation of the space of polynomials (explicitly, the span of the Ψπ ) which is isomorphic to the dual of the representation on the link patterns. By dual we mean, defined by composing the anti-isomorphism that sends the ei to themselves and ρ to ρ−1 (but reverses the order of the products) and transposition. The role of the anti-isomorphism is most evident when we try to compose equations (4.7,4.8) and find that êi êj Ψ = ej ei Ψ (in fact, it is often advocated to make one of the sets of operators act on the right, though we find the notation too cumbersome to use here). So the search for polynomial solutions of (4.2,4.3) is equivalent to finding certain irreducible sub-representations of the action of affine Hecke on the space of polynomials. Remark: the direct relation between qKZ and representations of an appropriate affine algebra only works for the An series of algebras i.e. affine Hecke. For more complicated situations such as the Birman–Wenzel–Murakami (BWM) algebra, it does not work so well because one cannot separate the two different actions [88]. 4.2. Construction of the solution. On general grounds, we only expect polynomial solution for integer values of the level. Here we shall only need a solution at level 1, that is s = q 6 . 4.2.1. qKZ as a triangular system. We shall build this solution in several steps. First, we use a “nice” property of our basis of link patterns, that is, the fact that (4.6) can be written as a triangular linear system in the components of ΨL . This requires to define an order on link patterns, which is most conveniently described as follows. Draw once again link patterns as pairings of points on a line and consider the operation described on Fig. 21. It gives a bijection between link patterns of size 2n and Young diagrams inside the staircase diagram 1n = (n − 1, n − 2, . . . , 1). Then the order is inclusion of Young diagrams. The smallest element, corresponding to the empty Young diagram, is denoted by 0n ; it connects i and 2n + 1 − i (note that it is one of the link patterns with all pairings parallel, which correspond to the smallest probability 1/An in the loop model). Consider now the exchange equation (4.6) and write it in components; we find two possibilities: 44 • i and i + 1 are not paired. Then we find that (4.6) only involves Ψπ , and implies that q zi − q −1 zi+1 divides Ψπ , and furthermore Ψπ /(q zi − q −1 zi+1 ) is symmetric in the exchange of zi and zi+1 . • i and i + 1 are paired. Then X (q zi − q −1 zi+1 )∂i Ψπ = Ψπ ′ π ′ 6=π,ei ·π ′ =π that is it involves the sum over preimages of π by ei viewed as acting on the set of link patterns. It turns out there are two types of preimages of a given π: in terms of Young diagrams, there is the Young diagram obtained from π by adding one box at i, i + 1 (which is always possible unless π is the largest element); and there are other Young diagrams that are included in π. So we can write the equation X Ψπ+one box at (i,i+1) = (q zi − q −1 zi+1 )∂i Ψπ − Ψπ ′ (4.9) π ′ ⊂π ei ·π ′ =π which has the desired triangular structure and allows to build the Ψπ one by one by adding boxes to the corresponding Young diagram. However there is no equation for Ψ0n . In fact this triangular system can be explicitly solved [54] (see also [19]) in the sense that every Ψπ can be written as a series of operators acting on Ψ0n . We shall not use this explicit solution here. From the discussion above, we find that all we need is to fix Ψ0n . We use the following simple observation, which generalizes the first case in the dichotomy above: Q If there are no pairings between points i, i+1, . . . , j in π, then i≤p<q≤j (q zp −q −1 zq ) divides Ψπ . (the proof is by induction on j − i). We now make the “minimality” assumption that in the case of 0n , these factors form all of Ψ0n , so that we find Y Y (4.10) Ψ0n = (q zi − q −1 zj ) (q −1 zj − q zi ) 1≤i<j≤n n+1≤i<j≤2n where we recall that the system size is L = 2n. If a solution of the qKZ system with such a Ψ0n exists, then it will be the only solution of degree n(n − 1) (up to multiplication by a constant). It remains however a non-trivial fact that with such a choice of Ψ0n , the construction above is consistent (i.e. independent of the order in which one adds boxes), and that (4.3) is satisfied, with s = q 6 . This is the subject of the next section. Example: we find in size L = 6 q 2 z1 − z2 q 2 z1 − z3 q 2 z2 − z3 q 2 z4 − z5 q 2 z4 − z6 q 2 z5 − z6 5 Ψ6 4 = − q6 1 3 2 Ψ6 5 1 2 Ψ6 5 4 1 q 2 z3 − z4 q 2 z5 − z6 (z1 z2 z3 q 8 + z1 z2 z4 q 8 − z1 z3 z4 q 6 − z2 z3 z4 q 6 4 = q7 3 − z1 z2 z5 q 6 − z1 z2 z6 q 6 + z3 z4 z5 q 2 + z3 z4 z6 q 2 + z1 z5 z6 q 2 + z2 z5 z6 q 2 − z3 z5 z6 − z4 z5 z6 ) q 2 z1 − z2 =− q 2 z2 − z3 q 2 z2 − z4 q 2 z3 − z4 q8 q 4 z1 − z5 3 2 45 q 4 z1 − z6 q 2 z5 − z6 Ψ6 5 4 1 q 2 z3 − z4 q 2 z3 − z5 q8 q 2 z4 − z5 q 4 z1 − z6 q 4 z2 − z6 3 2 Ψ6 =− q 2 z1 − z2 5 1 2 q 2 z4 − z5 q 4 z1 − z6 (z1 z2 z3 q 10 − z2 z3 z4 q 6 − z2 z3 z5 q 6 − z1 z4 z5 q 6 4 = q9 3 − z1 z2 z6 q 6 − z1 z3 z6 q 6 + z2 z4 z5 q 4 + z3 z4 z5 q 4 + z2 z3 z6 q 4 + z1 z4 z6 q 4 + z1 z5 z6 q 4 − z4 z5 z6 ) q 2 z2 − z3 4.2.2. Consistency and Jucys–Murphy elements. Let us go back to the qKZ system (4.7,4.8). In order to show that this system of equations is consistent, one must simply simply check that one obtains the same representation of affine Hecke on both sides (up to duality). Thus, we shall now check that the center has the same value in both representations. This will fix κ and more importantly s. Let us first define a new set of elements of the affine Hecke algebra (Jucys–Murphy elements), namely −1 yi = ti · · · tL−1 ρ−1 t−1 1 · · · ti−1 i = 1, . . . , L For this section we shall use “knot-theoretic” drawings to explain the equations we write. It is easy to see that if the ti = q −1/2 ei + q 1/2 are crossings in the knot theoretic sense (this definition being itself the skein relation for the Jones polynomial), then the yi have an equally simple description: ti = i i+1 yi = i where the second picture is embedded in a strip that is periodic in the horizontal direction. Then it becomes graphically clear that the yi commute: = yi yj = i j = yj yi i j One can also show that the symmetric polynomials of the yi′ = q −i yi form the center of the affine Hecke algebra.5 Thus, assuming that we have an irreducible representation such that we can diagonalize the action of the yi′ , the (unordered) set of their L eigenvalues is independent of the eigenvector and characterizes the representation. Let us now apply this to our two representations. First, let us make the yi act on link patterns. They have an obvious common eigenvector: the link pattern 1n which pairs neighbors 2i − 1 and 5This extra factor of q i comes from our normalization of the t . In the Hecke algebra they have no preferred i normalization, and one usually chooses it to make this factor disappear, but in Temperley–Lieb there is one (which makes the skein relation rotationally invariant), that we have chosen here, and it is unfortunately different. 46 2i, i = 1, . . . , n. Indeed, we find: 7 = −q −3/2 yi 1n = 1 6 8 5 1 4 2 3 i i.e. that the action of yi is equivalent to a Reidemeister move I (with negative orientation for i even and positive orientation for i odd). A move I multiplies the Jones polynomial by −q −3/2 (this amounts to the calculation ti ei = −q −3/2 ei ), so that we obtain i yi 1n = −q (−1) 3/2 1n (4.11) −1 −1/2 ê + q −1/2 . On the polynomial side, define similarly ŷi = t̂−1 i i−1 · · · t̂1 ρ̂t̂L−1 · · · t̂i , where t̂i = q One can show that there is an order on monomials (see [89]) such that the operators ŷi are upper triangular. Thus it suffices to evaluate their diagonal elements in the basis on monomials. The Q λi action of ŷi is most easily computed on monomials L i=1 zi with non-increasing powers λi+1 ≤ λi . In this case we find (4.12) ŷi L Y i=1 1 ziλi = −κsλi q 3(n+ 2 −i) L Y ziλi + lower terms i=1 Comparing the two expressions (4.11) and (4.12), we find that s = q 6 and λ = (λi ) is, up to a global shift, equal to λ(n) = (n − 1, n − 1, . . . , 1, 1, 0, 0). This is the minimal degree solution that we shall consider – the degree being n(n − 1), as assumed earlier. We also deduce κ = q −3(n−1) . The latter computation looks very similar to a computation in nonsymmetric Macdonald polynomials [73]. This is no coincidence. In fact the eigenvectors of ŷi are exactly nonsymmetric Q λi (π) Macdonald polynomials, see [50]. If we restrict ourselves to leading terms L which are in i=1 zi bijection with link patterns by labelling closings/openings in decreasing order, e.g. π= 1 2 3 4 5 6 7 7−→ 8 λ(π) = (3, 2, 3, 1, 2, 1, 0, 0) then their span is exactly the span of the Ψπ , even though they are distinct from the Ψπ – there is a triangular change of basis, so that only Ψ0n is itself a nonsymmetric Macdonald polynomial. 4.2.3. Wheel condition. In [89], a particularly attractive characterization of the span of the Ψπ (z1 , . . . , zL ), π ∈ P2n , is obtained. It will give an independent check of some of the results of the last section, as we shall see. Consider the following subspace of homogeneous polynomials: (4.13) Mn = {f (z1 , . . . , z2n ), degzi f ≤ n − 1 : f (. . . , z, . . . , q 2 z, . . . , q 4 z, . . .) = 0} that is polynomials that vanish as soon as an ordered triplet of variables form the sequence 1, q 2 , q 4 . This vanishing condition is a so-called wheel condition, see [49] for more general ones. We have also imposed a bound of n − 1 on the degree in each variable, which is a little bit more restrictive 47 than what is really needed: imposing on the (total) degree deg f ≤ n(n − 1) would suffice, but this condition is less convenient for our purposes.6 It is easy to check that Ψ0n ∈ Mn ; more interestingly, the action of τ − êi = −(q zi − q −1 zi+1 )∂i preserves Mn (check the vanishing property by discussing separately the three cases: (i) neither i or i + 1 are in the triplet: trivial; (ii) both i and i + 1 are in the triplet: then the prefactor vanishes; and (iii) one of them is: then use the property for the original triplet and the one with i replaced with i + 1). Since the Ψπ can be built from Ψ0n by action of the τ − êi , one concludes that they are all in Mn .7 In fact, note that the action of ρ̂ also leaves Mn invariant on condition that s = q 6 , so that Mn is a representation of the affine Hecke algebra. We have an inclusion of the span of the Ψπ inside Mn . In order to prove the equality, we shall state certain useful properties. 4.2.4. Recurrence relation and specializations. First, we need the following simple observation. According to the dichotomy of section 4.2.1, if π is a link pattern such that the points i and i + 1 are unconnected, then q zi − q −1 zi+1 divides Ψπ , so that Ψπ (zi+1 = q 2 zi ) = 0. But if i and i + 1 are connected (we shall call such a pairing of neighbors a “little arch”), then we can use the wheel condition to say that Ψπ (zi+1 = q 2 zi ) vanishes when zj = q 4 zi , j > i + 1, or zj = q −2 zi , j < i, so Q Q2n 2 −4 that it is of the form i−1 j=1 (zi − q zj ) j=i+1 (zi − q zj )Φ where by degree consideration Φ can only depend on the zj with j 6= i, i + 1. It is already clear that Φ is in Mn−1 and so is a linear combination of the entries of ΨL−2 ; but there is better. Call ϕi the mapping from PL−2 to PL which inserts an extra little arch at (i, i + 1). Then we claim that 0 π 6∈ Im ϕi 2n i−1 Y Y (q 3 zi − q −1 zj ) (zi − q 2 zj ) (4.14) Ψπ (. . . , zi+1 = q 2 zi , . . .) = q −(n−1) π = ϕi π ′ j=i+2 j=1 Ψπ′ (z1 , . . . , zi−1 , zi+2 , . . . , zL ) A direct proof of this formula is particularly tedious. Let us instead use the following trick. Consider the following cyclic invariance property: the formulae (4.14) for different values of i = 1, . . . , L − 1 can be deduced from each other by application of (4.3). In fact, one can even extend this way (4.14) to i = L; being a little careful of the factors of s = q 6 that appear here and there, we obtain: (4.15) ΨϕL (π′ ) (z1 = q −4 zL , . . . , zL ) = q −4(n−1) 2n−1 Y j=2 (zL − q 2 zj )Ψπ′ (z2 , . . . , zL−1 ) where ϕL adds an arch between L and 1 (on top of all other arches in the half-plane picture of link patterns). So, to prove (4.14), it is enough to prove (4.15). We shall use the construction of section 4.2.1. Note that the mapping ϕL is particularly natural via the bijection to Young diagrams, since it becomes the embedding of the set of diagrams inside 1n−1 to that of diagrams inside 1n . In particular 0n−1 is sent to 0n ; so the first check is compatibility with (4.10), i.e. that when π = 0n in (4.15) we obtain in the r.h.s. the correct prefactors times Ψ0n−1 . The second check is that the formula (4.9) which allows to compute the components of ΨL , when 6In fact, one can show that having a partial degree greater than n − 1 while retaining a total degree of n(n − 1) would lead to a contradiction with our representation of affine Hecke using (4.12). 7A more visual proof of this result is given in [108], where it is noted that since (i, i + 1) and (i + 1, i + 2) cannot be both pairings, ΨL (zi , zi+1 = q 2 zi+1 , zi+2 = q 4 zi ) = 0 and then use (4.2) to permute the arguments and get the general wheel condition. 48 restricted to diagrams inside 1n−1 , produces the components of ΨL−2 . This is rather obvious graphically: the action of ei is compatible with the natural inclusion of diagrams (it can be defined directly in terms of Young diagrams). Furthermore, in order to remain inside 1n−1 , one must never add boxes at (1, 2) or (L − 1, L). So we never affect the variables z1 , zL and up to a shift of indices i → i − 1, this is just the procedure in size L − 2. This is enough to characterize entirely the r.h.s. of (4.15). Next, we need two more closely related properties: • If f ∈ Mn , then (4.16) f (q −ǫ1 , . . . , q −ǫ2n ) = 0 ∀(ǫi ) Dyck paths increments ⇒ f =0 P Pj (a Dyck path increment is a sequence (ǫi )i=1,...,2n of ±1 such that 2n i=1 ǫi = 0 and i=1 ǫi ≥ 0 for all j ≤ 2n). So a polynomial in Mn is entirely determined by its values at these cn specializations. • We have the specializations: (4.17) ( (−1)n(n−1)/2 (q − q −1 )n(n−1) τ |π| if ǫi = sign(π(i) − i), i = 1, . . . , 2n −ǫ1 −ǫ2n Ψπ (q , . . . , q )= 0 otherwise (|π| is the number of boxes in the correspondence with Young diagrams of Fig. 21). In other words, these values are the entries in the basis of the Ψπ . The first point is shown explicitly in appendix C of [34] and we shall not reproduce the proof here; the second one is simply an application of the recurrence formula (4.14), by removing little arches one by one (see also [28]). 4.2.5. Wheel condition continued. Using the properties of the previous section, we can now conclude that dim Mn ≤ cn and that the Ψπ are independent polynomials, so that Mn is the span of the Ψπ . One could use more representation-theoretic arguments to prove the equality of Mn and of the span of the Ψπ , for example based on the fact that Mn , as a representation of the affine Hecke algebra, is irreducible for generic q (in fact, it corresponds to the rectangular Young diagram with 2 rows and n columns), but this is not our philosophy here. The wheel condition has the following physical interpretation (borrowed once again from [89]). Consider the case q = ±1. Then the qKZ equation (4.5) becomes the ordinary KZ equation at level 1. It is most conveniently expressed in the spin basis: L X σi,j + 1 ∂ Ψ= Ψ 3 ∂zi zi − zj j=1 j6=i where σi,j exchanges spins i and j. The KZ equation is now in its usual form for correlation d , except for a trivial change of σi,j − 1/2 to σi,j + 1, which is the same as functions in sl(2) 1 Q Ψ → Ψ i<j (zi − zj )1/2 (which is needed since we want a polynomial solution and not the usual one involving square roots). The solution is a product of two Vandermonde determinants: #{i<j:αi =−,αj =+} (4.18) Ψspin ∆(zi )αi =+ ∆(zi )αi =− α1 ,...,αL (z1 , . . . , zL ) = (−1) (α1 , . . . , αL ) ∈ {+, −}L 49 This is simply the ground state wave function of two species of free fermions ψ± , i.e. with notations similar to section 1 Ψspin α1 ,...,α2n (z1 , . . . , z2n ) = h0, 0| ψα1 (z1 ) · · · ψαn (zn ) |n, ni In this language the wheel condition is just the Pauli exclusion principle (if three fermions sit at the same spot, at least two of them must be of the same kind!). Of course the q = ±1 case is trivial in the sense that starting from say Ψ0n , one can apply the exchange equation (4.2) which simply says in this limit, using Ři = −σi,i+1 , that exchanging variables is the same as exchanging spins (up to a fermionic minus sign), which immediately results in (4.18). For a general value of q, one has a q-deformed version of this system of free fermions in which the vanishing condition of the wave function occurs not at coinciding points but when ratios of coordinates are ordered successive powers of q 2 . The usual fermionic statistic is turned into a non-trivial braid group statistic. 4.3. Connection to the loop model. In general, the two problems of diagonalizing the transfer matrix of the CPL loop model and finding solutions of the qKZ system are unrelated. However there is exactly one value of q where a solution of qKZ does in fact provide an eigenvector of the transfer matrix (with periodic boundary conditions). This is when the parameter s = 1, which here occurs when q = e2πi/3 (other sixth roots of unity are possible but they are either trivial or give the same result as the one we picked). In this case note that (4.3) becomes a simple rotational invariance condition. Furthermore the real qKZ equation (4.5) becomes an eigenvector equation for the scattering matrices: Si (z1 , . . . , zL )ΨL (z1 , . . . , zL ) = ΨL (z1 , . . . , zL ) These scattering matrices do not involve any extra shifts of the spectral parameters, and as is well-known in Bethe Ansatz, are just specializations of the inhomogeneous transfer matrix. Indeed if we define TL (z; z1 , . . . , zL ) to be simply ··· zL z1 z2 (with periodic boundary conditions), where the box represents the R-matrix evaluated at the ratio of vertical and horizontal spectral parameters, then observe that Si (z1 , . . . , zL ) = TL (zi ; z1 , . . . , zL ). By a Lagrange interpolation argument, we conclude that TL = z TL (z; z1 , . . . , zL )ΨL (z1 , . . . , zL ) = ΨL (z1 , . . . , zL ) i.e. ΨL is up to normalization the steady state of the inhomogeneous Markov process defined by TL (z; z1 , . . . , zL ). In order to recover the original homogeneous Markov process, one simply sets all zi = 1. Note that the reasoning above is valid for any solution of qKZ, not just the one that was discussed in section 4.2.1. But there is no point in considering higher degree polynomial solutions since at q = e2πi/3 they will simply be the minimal degree of section 4.2.1 times a symmetric polynomial. We now show how to apply the qKZ technology to prove some of the statements formulated in section 3. 4.3.1. Proof of the sum rule. As a simple application of the above formalism, we explain how to recover the “sum rule” at q = e2πi/3 . Let us explain what we mean by that. We start by noting that the normalization of Ψ is fixed once the value of Ψ0n is specified; in particular, if all zi = 1, using (4.10) we find Ψ0n = 3n(n−1)/2 , to be compared with the (conjectured) probability 1/An associated to 0n , cf section 3.3.3. In other words, with this normalization, one should have 50 P Ψπ = 3n(n−1)/2 An . This is what we are going to show now. P In fact, following [23], we are going to show more generally the inhomogeneous sum rule that π∈P2n Ψπ (z1 , . . . , z2n ) is equal to the Schur function sλ(n) (z1 , . . . , z2n ), where λ(n) is defined in (2.16), that is up to a constant the partition function of the six-vertex model with DWBC at q = eiπ/3 . Instead of the inductive method of [23], i.e. the comparison of the recurrence relations of sections 4.2.4 and 2.5.2, we shall here proceed directly. π∈P2n Define the covector v with entries vπ = 1, π ∈ P2n . The stochasticity property of a matrix is the fact that v is left eigenvector with eigenvalue 1; and it is satisfied at τ = 1 by all the ei and henceforth by Ři (z): q = e2πi/3 P Applying this identity to (4.2), we immediately conclude that v · Ψ = π∈P2n Ψπ is a symmetric polynomial of its arguments z1 , . . . , zL . P π∈P2n Ψπ (z1 , . . . , z2n ) is a symmetric polynomial of degree n(n − 1) which satisfies the wheel condition of (4.13). We claim that this defines it uniquely up to normalization (see [108] for a similar claim in a more general fused model). The simplest proof is to use once again property (4.16) i.e. that polynomials in Mn are characterized by the specializations (q −ǫ1 , . . . , q −ǫ2n ). But for a symmetric polynomial, all these specializations reduce to only one. This proves the claim. vei = v ⇒ v Ři (z) = v i = 1, . . . , L − 1, Next, one notes that sλ(n) (z1 , . . . , z2n ) also satisfies these properties. The degree bound is clear from the definition (1.24) of Schur functions; as to the wheel condition, it follows from formula (1.23) by noting that if zk = q 2 zj = q 4 zi , the sub-matrix of the matrix of the numerator with columns {i, j, k} is of rank 2 and thus the determinant of the whole matrix vanishes. Finally, to fix the normalization constant, note that thanks to (4.17), all components of Ψ vanish at (z1 , . . . , z2n ) = (q −1 , . . . , q −1 , q, . . . , q) except Ψ0n (q −1 , . . . , q −1 , q, . . . , q) = 3n(n−1)/2 ; on the other hand, using (2.17), we find the same value sλ(n) (q −1 , . . . , q −1 , q, . . . , q) = 3n(n−1)/2 . We conclude that X Ψπ (z1 , . . . , z2n ) = sλ(n) (z1 , . . . , z2n ) π∈P2n Remark: from the representation-theoretic point of view, v is an invariant element (under affine P Hecke action) in (CP2n )⋆ ; and π∈P2n Ψπ is its image in C[z1 , . . . , zL ]. This shows that at q = e2πi/3 , the representation of affine Hecke on CP2n (resp. the span of the Ψπ ) is not irreducible (though it remains indecomposable), since it has a codimension one (resp. dimension one) stable subspace. 4.3.2. Case of few little arches. Note a remarkable property of recurrence relation (4.14): it can only decrease the number of little arches! So, if one considers the subset of link patterns with a given maximum number of little arches, that subset is closed under these relations. The link patterns that possess only two little arches (which is the minimum) are the link pattern 0n and its rotations, i.e. the n link patterns for which all the pairings are parallel and for which we know explicitly the components: they are given by the action of ρ̂ (cyclic rotation of variables plus shift of s) on Ψ0n which is (4.10). These are the smallest components in the homogeneous limit. They correspond via the Razumov– Stroganov conjecture to a single FPL. As such they are clearly the “easiest” components. It is natural to look at what happens when one considers more little arches. Link patterns with three little arches are of the form of Fig. 22. In fact, the case of three little arches was first investigated in [29] on the other side of the Razumov–Stroganov conjecture: there, the problem of the enumeration of FPLs with such a connectivity was solved, with the remarkable result that the 51 b a a b c γa+b α1 q2 γ1 q2 c αb+c q 2 β1 q 2 βa+c Figure 22. Link patterns with 3 arches. number of FPLs with link pattern (a, b, c) is simply equal to the number of plane partitions inside an a × b × c box, that is given by the MacMahon formula (1.27). The proof is bijective and we shall not reproduce it here. Let us just comment on it. The crux of the bijection is the observation, known as “de Gier’s lemma”, that in an FPL with given connectivity many edges are fixed by the simple requirement of the connectivity of the endpoints. Removing all these fixed edges we obtain a simpler enumeration problem, typically of plane partitions. Now the little arches play a key role in the sense that they are the ones that limit application of de Gier’s lemma. In other words, the more little arches there are, the fewer fixed edges in FPLs, and therefore the more complicated the enumeration is. Once again, the number of little arches is a measure of complexity. Closing this philosophical parenthesis, let us go back to the qKZ equation and try to compute the entries Ψa,b,c of ΨL corresponding to link patterns with three series of a, b, c nested arches. This is performed in [107] for q = e2πi/3 and we generalize it here to arbitrary q. Up to rotation and use of (4.3), one can always assume that one of the little arches, say the one that is part of the series of a nested arches, is between L and 1. Let us further rename the L variables zi as follows: they become (q −2 γ1 , . . . , q −2 γa+b , α1 , . . . , αb+c , q 2 β1 , . . . , q 2 βa+c ), see Fig. 22. Then according to the results of section 4.2.1, Y Y Y Ψa,b,c = (q αi − q −1 αj ) (q βi − q −1 βj ) (q γi − q −1 γj ) 1≤i<j≤b+c 1≤i<j≤a+c 1≤i<j≤a+b q −b(a+b−1)+c(a+3b−1)+c2 (−1) (a+b)(a+b−1)+c(b+(c−1) 2 Φa,b,c where Φa,b,c is a polynomial that is symmetric in the {αi }, in the {βi } and in the {γi }. The powers of q and sign have been chosen carefully so that when we rewrite recurrence relation (4.14) for zi = αb+c , zi+1 = q 2 β1 in terms of Φa,b,c , everything cancels out and we are left with (4.19) Φa,b,c |β1 =αb+c = a+b Y (αb+c − γk )Φa,b,c−1 k=1 where the arguments β1 , αb+c are missing from Φa,b,c−1 . Now since Φa,b,c is symmetric in the αi , this relation provides Φa,b,c at b + c values of β1 ; and it is not hard to check that Φa,b,c is of degree b in β1 . Thus it is entirely determined by these values (and can for example be obtained by Lagrange interpolation). To complete the (implicit) calculation of Φa,b,c one must provide an initial condition. At c = 0, the link pattern (a, b, 0) is nothing but the base link pattern 0n , and in this case we find (4.20) Φa,b,0 = b a Y Y (αj − βi ) i=1 j=1 Some explicit formulae for Φa,b,c are given in [107]. In fact they are “triple Schur functions” in the sense of multi-Schur functions of [63]. 52 γ4 γ1 b γ3 γ2 γ5 β1 γ6 c α1 a α2 α3 γ4 β2 β3 α7 β4 α6 β5 α5 α4 γ1 b γ3 γ2 γ5 β1 γ6 c a (a) α1 α2 α3 β2 β3 α7 β4 α6 β5 α5 α4 (b) Figure 23. The three sets of spectral parameters associated to lozenge tilings. From the discussion above, it is natural to try to put non-trivial Boltzmann weights on lozenge tilings of a hexagon of sides a, b, c in such a way that their partition function inside an a × b × c box coincides with the inhomogeneous component Φa,b,c . Such a model is introduced in [107]. We point it out here because we believe this model is of some interest and might deserve further study. The spectral parameters live on the medial lattice, which is the Kagome lattice, see Fig. 23(a). Each lozenge has a weight which is the difference of spectral parameters crossing at the center of the lozenge, with the sign convention α − β, β − γ, α − γ.8 One can show that the integrability of the underlying model on the Kagome lattice implies that the partition function is a symmetric function of the {αi }, of the {βi } and of the {γi }. We choose not to do so here and refer instead to the appendix of [107] for a “manual” proof of symmetry. Finally, it is easy to see that the partition function of lozenge tilings with such weights satisfies the recurrence relation (4.19), as illustrated on Fig. 23(b); and that when c = 0, there is a unique tiling of the a × b parallelogram, resulting in (4.20). Note the similarity with the properties of the partition function of the six-vertex model with DWBC (symmetry and recurrence) found by Korepin. Thus, Φa,b,c and this partition function coincide. Finally, if αi = 1, βi = q −2 , γi = q 2 , using the fact that the number of lozenges of each orientation is fixed and equal to ab, bc, ca, we find: Φa,b,c = (1 − q −2 )ab (q −2 − q 2 )bc (1 − q 2 )ca Na,b,c = (q − q −1 )ab+bc+ca q a(c−b) (−1)ca Na,b,c q = e2πi/3 2 Ψa,b,c = q −(a+c)(a+c−1)+(a+b)(a+b−1)−b(a+b−1)+c(a+3b−1)+c (−1) (q − q −1 ) (b+c)(b+c−1) + (c+a)(c+a−1) + (a+b)(a+b−1) 2 2 2 = (q − q −1 )n(n−1) (−1) n(n−1) 2 (a+b)(a+b−1)+c(b+(c−1) 2 Φa,b,c q = e2πi/3 Na,b,c We conclude that Ψa,b,c /Ψ0n = Na,b,c , as expected. The case of four little arches is similar to that of three arches and is sketched in [107]. On the one hand, the recurrence relations are once again enough to determine all Ψπ uniquely. On the other 8This is only a convention because the total number of lozenges of each orientation is conserved. The choice that we made does not respect the Z3 symmetry of the model but turns out to be convenient. 53 hand, the enumeration of FPL configurations with four little arches is doable (after appropriate use of the so-called Wieland rotation [100]) and can be once again reduced to a lozenge tiling enumeration [30]. So the same strategy applies. Starting with five little arches, however, it fails on both sides: the recurrence relations are not sufficient to determine the corresponding components Ψπ , and the enumeration of FPLs becomes non-trivial due to an insufficient number of fixed edges. 4.4. Integral formulae. We now want to show that, using the formalism of the qKZ equation allows to prove the properties discussed in 3.3.3, as well as to reconnect the three models that we have found in which the same numbers An appear. A particular useful tool to exploit these solutions of qKZ has been introduced in [28, 92]: it consists in writing integral formulae for them. 4.4.1. Integral formulae in the spin basis. There are various types of known integral formulae for solutions of the quantum Knizhnik–Zamolodchikov equation. Here we are only interested in a very specific one, which only exists at level 1 and is obtained by (q-)bosonization, see [45]. The following formula is valid in even size L = 2n: Y −1 n (q zi − q −1 zj ) (4.21) Ψspin a0 ,...,an−1 (z1 , . . . , zL ) = (q − q ) 1≤i<j≤L I Q I n−1 −1 Y wℓ dwℓ 0≤ℓ<m≤n−1 (wm − wℓ )(q wℓ − q wm ) ··· Q Qn−1 Q −1 2πi a ≤i≤L (q wℓ − q zi ) ℓ=0 1≤i≤a (wℓ − zi ) ℓ=0 ℓ ℓ where the contours surround the zi counterclockwise, but not the q −2 zi . The reader must be warned that we have changed the labelling of the entries of Ψ: the index, instead of being a sequence of L spins, is the ordered sequence of locations of plus spins. In other words the correspondence is (α1 , . . . , αL ) 7→ {a0 < · · · < an−1 } = {i : αi = +}. For the case of odd size see [92]. 4.4.2. The partial change of basis. Next we would like to write similar expressions for the entries of Ψ in the basis of link patterns. Unfortunately this would require to invert the change of basis from link patterns to spins that was described in section 3.2.4; and there is no simple explicit formula for the inverse (this is the famous problem of the computation of Kazhdan–Lusztig polynomials [52]; in this case there is a combinatorial formula [64] but it is not obvious how to combine it with integral formulae). Instead we shall do here something more modest: we introduce an intermediate basis between spins and link patterns. We skip here the details, which can be found in [28] (see in particular the appendices) and [27]. The bottom line is the introduction of a slight improvement of formula (4.21): (4.22) Ψa0 ,...,an−1 (z1 , . . . , zL ) = Y (q zi − q −1 zj ) 1≤i<j≤L I Q I n−1 −1 Y dwℓ 0≤ℓ<m≤n−1 (wm − wℓ )(q wℓ − q wm ) ··· Q Qn−1 Q 2πi ℓ=0 1≤i≤a (wℓ − zi ) a <i≤L (q wℓ − q −1 zi ) ℓ=0 ℓ ℓ (note the subtle modification: the factors of wℓ disappeared and an inequality became strict in the denominator). Here (a0 , . . . , an−1 ) is an arbitrary non-decreasing sequence of integers. The 54 connection with the entries of Ψ in the bases of link patterns and spins is as follows: P X Ψspin (−q)− i εi Ψa0 −ε0 ,...,an−1 −εn−1 a0 ,...,an−1 = ε0 ,...,εn−1 ∈{0,1} Ψa0 ,...,an−1 = X π∈P2n Y ! U#{ℓ: i≤aℓ <j}−(j−i+1)/2 Ψπ i,j paired in π where Uk is the Chebyshev polynomial of the second kind evaluated at −τ : Uk = q k+1 −q −(k+1) . q−q −1 Note an interesting property of the second change of basis, which is that the coefficients are polynomials in τ with integer coefficients. A second related property appears when one takes the homogeneous limit zi = 1 in the integral formula (4.22). It is convenient to send the two poles (1 and q −2 ) to zero and infinity respectively by the homographic transformation uℓ = (wℓ − 1)/(q wℓ − q −1 ), so that: I I n−1 Y duℓ Y Ψa0 ,...,an−1 (1, . . . , 1) = · · · (um − uℓ )(1 + τ um + uℓ um ) aℓ Ψ0n 2πi uℓ ℓ=0 0≤ℓ<m≤n−1 where τ = −q − q −1 , and the contours surround zero. This can be rewritten Y Ψa0 ,...,an−1 (1, . . . , 1) = (um − uℓ )(1 + τ um + uℓ um )Qn−1 uai −1 i=0 i Ψ0n 0≤ℓ<m≤n−1 As shown in [27], one can find a subset of sequences (a0 , . . . , an−1 ) such that the matrix of change of basis to link patterns is upper triangular with 1’s on the diagonal. Combining the two properties above, we reach the important conclusion that, normalized so that Ψ0n = 1, the entries Ψπ of the level 1 polynomial solution of qKZ in the homogeneous limit zi = 1 are polynomials of τ with integer coefficients. In fact, these coefficients are non-negative, but no combinatorial meaning is known for them. We shall however see in the next section that we can interpret coefficients of certain partial sums of entries in terms of plane partitions. Example: in size L = 6, we find the entries of Ψ to be, with the same ordering of link patterns as in sections 3.2.3 and 4.2.1, the polynomials 1, 2τ , τ 2 , τ 2 , τ + τ 3 . 4.4.3. Sum rule and largest component. The integral formulae above allow to compute various linear combinations of the Ψπ . Here, we concentrate on a single quantity P of particular interest (which is considered in [28, 20]). Consider the linear combination Fn (t) = π∈P2n cπ (t)Ψπ (1, . . . , 1), with the normalization Ψ0n = 1, and where the coefficient cπ (t) is most simply expressed in terms of the corresponding Young diagram by assigning a weight of t or t−1 to the boxes of even/odd rows in the complement inside 1n , see Fig. 24. Three special cases are of particular interest: F (1) is just P the sum π∈P2n Ψπ of all link patterns; F (0) is just Ψ1n , one of the two largest components; and F (t)|tn−1 , the leading term of F (t), is just Ψρ(1n ) , the other largest component. P We then claimPthat π∈P2n cπ (t)Ψπ can be naturally expressed in the intermediate basis as P n i=1 ǫi Ψ t 1−ǫ0 ,3−ǫ1 ,...,2n+1−ǫn−1 (this is a simple calculation of the corresponding enǫ0 ,...,ǫn−1 ∈{0,1} tries of the matrix of change of basis), which results in an integral formula. In the homogeneous limit we obtain Fn (t) = n−1 Y (1 + t uℓ ) ℓ=0 Y 0≤ℓ<m≤n−1 (um − uℓ )(1 + τ um + uℓ um )Qn−1 u2i i=0 55 i t−1 t t t−1 t t Figure 24. Weight of a component expressed in terms of the associated Young diagram. In fact one can simply set u0 = 0 to get (4.23) Fn (t) = n−1 Y Y (1 + t uℓ )(1 + τ uℓ ) ℓ=1 (um − uℓ )(1 + τ um + uℓ um )Qn−1 u2i−1 i=1 1≤ℓ<m≤n−1 i which looks strikingly similar to the formula for the weighted enumeration TSSCPPs (1.29). There is however an important difference. Whereas (1.29) just contains a product of functions of one variable times an antisymmetric function of the ui (which, ultimately, comes from the free fermionic nature of the model), (4.23) does not possess any particular symmetry w.r.t. exchange of its variables, due to the factors 1 + τ um + uℓ um (which come from the q-deformed Vandermonde product in the original integral formula). One can however antisymmetrize formula (4.23), as conjectured in [28] and then subsequently proved in [103] and in a slightly stronger version in [34]. The identity, as formulated in [20], is: P σ If AS(f (u1 , . . . , un−1 )) = σ∈Sn−1 (−1) f (uσ(1) , . . . , uσ(n−1) ), and (· · · )≤0 means keeping only non-positive powers of a Laurent polynomial in the variables uℓ , then the following equality holds: n−1 Y Y Y −2ℓ+1 uℓ (1 + uℓ um + τ um ) (4.24) (1 − uℓ um ) AS ℓ=1 1≤ℓ≤m≤n−1 = AS n−1 Y −1 ℓ−1 u−ℓ ℓ (τ + uℓ ) ℓ=1 1≤ℓ<m≤n−1 ! = n−1 Y ℓ=1 u−1 ℓ ≤0 Y −1 −1 −1 (u−1 m − uℓ )(τ + uℓ + um ) 1≤ℓ<m≤n−1 Applying (4.24) to (4.23), we find (4.25) Fn (t) = Y 1≤i≤j≤n−1 n−1 uj − ui Y (1 + t ui )(1 + τ ui )i Qn−1 u2i−1 i=1 i 1 − ui uj i=1 If t = 1, this coincides exactly with (1.29). What we have found is a very non-trivial combinatorial interpretation of this polynomial solution of qKZ, which goes beyond the special value q = e2πi/3 : for generic q, the sum of components of this solution reproduces the weighted enumeration of TSSCPPs with weight τ = −q − q −1 , proving a conjecture formulated in [22]. Next, we deduce from (4.25) that the leading term of Fn (t) is Fn (t)tn−1 = Y 1≤i≤j≤n−1 n−1 uj − ui Y (1 + τ ui )i Qn−1 u2i−2 = Fn−1 (τ ) i=1 i 1 − ui uj i=1 where the last equality is obtained by setting u1 = 0 and relabelling i → i − 1. With a bit more work, one can also prove that Fn (0) = τ n−1 Fn−1 (1/τ ). 56 As a result, at τ = 1, Fn (1), the sum of components, is An , and Fn (t)|tn−1 = Fn (0), the largest components, are An−1 . 4.4.4. Refined enumeration. For general t, one can show that Fn (t) corresponds to a refined weighted enumeration of TSSCPPs, that is to adding extra boundary weights on the TSSCPPs. The details can be found in [28]. Remarkably enough, if one sets τ = 1, then Fn (t) also coincides with the refined enumeration of ASMs. For ASMs, the refinement is simple to explain and consists in giving a weight of ti−1 to an ASM whose 1 in the first row is at column i (which is essentially equivalent to keeping one spectral parameter free in the partition function of the six-vertex model with DWBC and sending the others to 1). This strange coincidence was observed in [92] and proved via rather indirect arguments (making use of the Bethe Ansatz calculations of [91]). One can push this idea further: in 1986, Mills, Robbins and Rumsey conjectured that the double refinements of TSSCPPs and ASMs coincide [78]. Again, the double refinement of ASMs is easy to explain, consisting in weights ti−1 un−j for an ASM with a 1 at (1, i) and at (n, j) (which is equivalent to keeping two say horizontal spectral parameters free). The double refinement of TSSCPPs can be formulated in multiple ways and we refer the reader to appendix A of [34] for details. And in fact, using slight generalizations of the integrals above, one can give a direct proof of this conjecture and turn it into a theorem, as is done in [34]. 5. Integrability and geometry The goal in this section is to reinterpret the families of polynomials that give rise to a solution of the qKZ equation in terms of geometric data. For simplicity we shall stick to cohomology (as opposed to K-theory), which means that we only consider rational solutions of the Yang–Baxter equation. In terms of the parameter q, it means taking an appropriate q → ±1 limit (it is actually more convenient to consider q → −1 with our sign conventions). We shall consider three examples in what follows. The first one (matrix Schubert varieties) is there as warming up, though it already contains many of the key ingredients. The second example (orbital varieties) will correspond to the q → −1 limit of the solution of qKZ of section 4. The third one (Brauer loop scheme) in fact contains the first two as special cases and is the most interesting one. But first we need to introduce some technology i.e. an algebraic analogue of equivariant cohomology for affine schemes with a linear group action: multidegrees [76]. Note that in this section we give almost no proofs and refer to the papers for details. 5.1. Multidegrees. Our group will always be a torus T = (C× )N . T acts linearly on a complex vector space W . To a closed T -invariant sub-scheme X ⊆ W we will assign a polynomial mdegW X ∈ Sym(T ∗ ) ∼ = Z[z1 , . . . , zN ] (here T ∗ is viewed as a lattice inside the dual of the Lie algebra of T ) called the multidegree of X. 5.1.1. Definition by induction. This assignment can be computed inductively using the following properties (as in [48]): 1. If X = W = {0}, then mdegW X = 1. 2. If the scheme X has top-dimensional P components Xi , where mi > 0 denotes the multiplicity of Xi in X, then mdegW X = i mi mdegW Xi . This lets one reduce from the case of schemes to the case of varieties (reduced irreducible schemes). 3. Assume X is a variety, and H is a T -invariant hyperplane in W . (a) If X 6⊂ H, then mdegW X = mdegH (X ∩ H). 57 (b) If X ⊂ H, then mdegW X = (mdegH X) · (the weight of T on W/H). One can readily see from these properties that mdegW X is homogeneous of degree codimW X, and is a positive sum of products of the weights of T on W . 5.1.2. Integral formula. We can also reformulate the multidegree as an integral (generalizing the idea that the degree of a projective variety is essentially its volume). To the torus T acting linearly on the complex vector space W is naturally associated a moment map µ, which is quadratic. With reasonable assumptions (i.e. that the multigrading associated to the torus action is positive), one can map generators of T ∗ to complex numbers in such a way that this quadratic form is positive. Then one can formally write R dx e−πµ(x) mdegW X = R X dx e−πµ(x) W where it is implied that on both sides, this evaluation map has been applied. More explicitly, suppose that (xi )i=1,...,n is a set of coordinates on W which are eigenvectors of the torus action, with weights αi ; then P R −π ni=1 αi |xi |2 dx e P mdegW X = R X −π ni=1 αi |xi |2 dx e W Z n Pn Y 2 dx e−π i=1 αi |xi | αi = i=1 X where the αi must be evaluated to positive real numbers for the integral to make sense. So, the multidegree is just a Gaussian integral. The subtlety comes from the fact that in general X is a non-trivial variety to integrate on (in particular, it will be singular at the torus fixed point 0, the critical point of the function in the exponential!). In the case where X is a linear subspace of W , given by equations xi1 = · · · = xik = 0, which is the only case where the integral is really an ordinary Gaussian integral, we compute immediately mdegW X = αi1 . . . αik , as expected. 5.2. Matrix Schubert varieties. Matrix Schubert varieties are a slight variation of the classical Schubert varieties, whose study started Schubert calculus. They are simpler objects to define, being affine schemes (i.e. given by a set of polynomial equations in a vector space). 5.2.1. Geometric description. Let g = gl(N, C), b+ = {upper triangular matrices} ⊂ g, b− = {lower triangular matrices} ⊂ g, and similarly we can define the groups G = GL(N, C) and B± = {invertible upper/lower triangular matrices} ⊂ G. The matrix Schubert variety Sσ associated to a permutation σ ∈ SN is a closed sub-variety of g defined by the equations: Sσ = {M ∈ g : rank M j ≤ rank σ T j , i, j = 1, . . . , N } (5.1) i i where M j is the sub-matrix above and to the left of (i, j), and σ T is the transpose of the permutation i matrix of σ. Explicitly, Sσ is defined by a set of polynomial equations of the form, determinant of a sub-matrix of M is zero, which express the rank conditions. One can also describe Sσ as a group orbit closure: (5.2) S σ = B− σ T B+ 58 Evidently, we have Sσ−1 = (Sσ )T . Less evidently, the codimension of Sσ is the inversion number |σ| of σ. Examples: • σ = 1: then rank σ T j = min(i, j) which is the maximum possible rank of a i × j matrix. So i there is no rank condition and Sσ = g. • σ = σ0 , the longest permutation, σ0 (i) = N + 1 − i: this time rank σ T j = max(0, i + j − N ) and we find that Sσ0 is a linear subspace Sσ0 = {M = (Mij )i,j=1,...,N i 0 : Mij = 0 ∀i, j, i + j ≤ N } = 0 . . . ⋆ ··· ⋆ .. . ⋆ • All the other matrix Schubert varieties are somewhere in between. For example, if σ = (4132), then 0 0 0 1 1 0 0 0 M11 = M12 = M13 = 0 σT = ⇒ S = M = (M ) : ij i,j=1,...,4 (4132) 0 0 1 0 M21 M32 − M22 M31 = 0 0 1 0 0 Note that among the multiple determinant vanishing conditions, one can extract in this case four that imply all the others. In general, there are simple graphical rules to determine which conditions to keep. There is a torus T = (C× )2N acting on g, by multiplication on the left and on the right by diagonal matrices. The corresponding generators of its dual are denoted by x1 , . . . , xN (left) and y1 , . . . , yN (right). In fact, since multiplication by a scalar does not see the distinction between left and right, the torus acting is really of dimension 2N − 1, and this amounts to saying that all multidegrees we shall consider only depend on differences xi − yj . The Sσ are T -invariant; let us define the double Schubert polynomials Ξσ to be their multidegrees: Ξσ = mdegg Sσ Note that Ξσ−1 (x1 , . . . , xN |y1 , . . . , yN ) = Ξσ (y1 , . . . , yN |x1 , . . . , xN ), so the two sets of variables play symmetric roles. 5.2.2. Pipedreams. In [32] (see also [56]), a combinatorial formula for the calculation of double Schubert polynomials was provided. It can be described as configurations of a simple statistical lattice model coined in [56] (reduced) “pipedreams”. The pipedreams are made of plaquettes, similarly to loop models of section 3. The two allowed and (where the second one should be considered as two lines actually plaquettes are crossing). Furthermore, pipedreams have a specific geometry: they live in a right-angled triangle, as shown on the examples below. On the hypotenuse, the plaquettes are forced to be of the first type (and only one half of them is represented). A pipedream is reduced iff no two lines cross more than once. Alternatively, one can think of putting a non-local weight of zero to “bubbles” formed by two lines crossing twice. We assign weights to reduced pipedreams as follows: a crossing at row/column (i, j) has weight xi − yj . We 59 can now state the formula for double Schubert polynomials: X Ξσ (x1 , . . . , xN |y1 , . . . , yN ) = (weight of pipedream) reduced pipedreams of size N such that point i on vertical axis is sent to point σ(i) on horizontal axis i=1,...,N A sketch of proof of this formula will be given in the next two sections. Examples: • σ = 1: then only one pipedream contributes, the one with no crossings, of the type 1 2 3 4 1 2 ⇒ Ξ1 = 1 3 4 • σ = σ0 : again, only one pipedream contributes, this time with only crossings: 1 2 3 4 1 2 ⇒ 3 Ξσ0 = Y i+j≤N (xi − yj ) 4 Each factor of xi − yj has the meaning of weight of the equation Mij = 0. • If σ = (4132), there are two (reduced) pipedreams: 1 2 3 4 1 1 1 2 2 3 3 4 4 2 3 4 ⇒ Ξ(4132) = (x1 − y1 )(x1 − y2 )(x1 − y3 )(x2 + x3 − y1 − y2 ) Again, we recognize in each factor the weight of one equation of S(4132) . In general, as long as Sσ is a complete intersection (as many defining equations as the codimension), Ξσ is just the product of weights of its equations (i.e. linear factors). 5.2.3. The nil-Hecke algebra. In [33, 32], double Schubert polynomials are related to a solution of the Yang–Baxter equation (YBE) based on the nil-Hecke algebra. We shall reformulate this connection here in terms of our exchange relation. Define the nil-Hecke algebra by generators ti , i = 1, . . . , N − 1, and relations t2i = 0 ti ti+1 ti = ti+1 ti ti+1 60 ti tj = tj ti |i − j| > 1 As a vector space, the nil-Hecke algebra is isomorphic to C[SN ] (the ti being identified with elementary transpositions). It has an obvious graphical depiction in which basis elements σ ∈ SN permute lines; the product is the usual product in SN except that when lines cross twice, the result is zero: 1 2 3 4 1 (2143)(1342) = 2 3 4 = = (2431) 1 1 2 3 2 3 4 4 1 2 3 4 =0 (1423)(3124) = 1 2 3 4 (the expressions from right to left correspond to the pictures from bottom to top). The inversion number of σ is graphically its number of crossings. Define Ři (u) = 1 + u ti which is the associated solution to the rational Yang–Baxter equation, that is with additive spectral parameters: Ři (u)Ři+1 (u + v)Ři (v) = Ři+1 (v)Ři (u + v)Ři+1 (u) , and the identity is 1 = Note that ti = , that is 45 degrees rotated versions of our plaquettes, cf a similar remark in section 4.1.1. As usual, our solution of YBE also satisfies the unitarity equation: Ři (u)Ři (−u) = 1 P Now consider the formal object Ξ = σ∈SN Ξσ σ viewed as an element of nil-Hecke. We then claim that the following formulae hold: (5.3) (5.4) Ξ(x1 , . . . , xN |y1 , . . . , yN )Ři (xi+1 − xi ) = Ξ(x1 , . . . , xi+1 , xi , . . . , xN |y1 , . . . , yN ) Ři (yi − yi+1 )Ξ(x1 , . . . , xN |y1 , . . . , yN ) = Ξ(x1 , . . . , xN |y1 , . . . , yi+1 , yi , . . . , yN ) which is nothing but exchange relations. Note that if we consider X as a function of the first set of variables x1 , . . . , xN , the nil-Hecke algebra must act on the right on itself (ultimately, this is because of the transposition in (5.1) and (5.2)). These formulae are obvious if one defines Ξ in terms of pipedreams, that is if one writes Ξ as [32] 1’ 2’ 3’ 4’ y1 y2 y3 y4 x1 1 Ξ= x2 2 x3 3 x4 4 1’ 2’ 3’ 4’ 1 2 3 4 = 61 where each plaquette is an R-matrix evaluated at yj − xi (row i, column j), and each pipedream is interpreted as an element of nil-Hecke according to the convention of endpoints shown next to the equality. The proof of (5.3) is then the usual argument (cf for example section 2.5.2) of applying the Ř matrix between rows i and i + 1, moving it across using YBE and checking that it disappears once it reaches the hypotenuse; and similarly for (5.4). To conform with the literature, we shall now focus on the first form (5.3). Writing it explicitly in components, we have to distinguish as usual cases. Let us denote as before si the permutation of variables xi and xi+1 . • If σ(i) < σ(i + 1) (the lines starting at (i, i + 1) do not cross): then σ is not the image of ti acting by right multiplication, so that we find (5.5) Ξσ = si Ξσ i.e. Ξσ is symmetric under exchange of xi and xi+1 . • If σ(i) > σ(i + 1) (the lines starting at (i, i + 1) cross): this time we find that Ξσ + (xi+1 − xi )Ξσ′ = si Ξσ , where σ ′ is the unique preimage of σ under right multiplication by ti . Thus, (5.6) Ξσ′ = ∂i Ξσ where ∂i = xi+11−xi (si − 1), as before. Note that σ ′ satisfies σ ′ (i) < σ ′ (i + 1), so that (5.5) can be deduced from (5.6) (the image of ∂i is symmetric in xi , xi+1 ). In (5.6), σ ′ has one fewer crossings than σ. It is easy to show this way that, starting from Ξσ0 = Q i+j≤N (xi − yj ), where σ0 is the permutation with the most crossings (the highest inversion number), one can compute any Ξσ by repeated use of (5.6). We now explain what the geometric meaning of (5.6) is. This will prove that the pipedreams do compute the multidegrees of matrix Schubert varieties. 5.2.4. The Bott–Samelson construction. Let Li be the subgroup of G consisting of matrices which are the identity everywhere except in the entries Mab with a, b ∈ {i, i + 1}, and B±,i = B± ∩ Li . We use the following lemma, which is similar to lemma 1 of [58] (see also lemma 8 of [59]): (throughout the lemma the sign ± is fixed) Let V be a left Li -module, and let X be a variety in V that is invariant under scaling and under the action of B±,i ⊂ Li . Define the map µ : (Li × X)/B±,i → V that sends classes of pairs (g, x) (where B±,i acts on the right on Li and on the left on X) to g · x. If µ is generically one-to-one, then mdegV Im µ = ∓∂i mdegV X If on the other hand X is Li -invariant, then ∂i mdegV X = 0 Here we apply the lemma with V = g, Li acting on it by left multiplication, and X = Sσ . According to (5.2), X is B−,i -invariant. Furthermore, if σ(i) > σ(i + 1), one can check that the map µ is generically one-to-one, and the image is precisely Sσ′ where σ = σ ′ ti . Hence (5.6) holds. On the contrary, if σ(i) < σ(i + 1), X is simply Li -invariant, and we conclude that (5.5) holds. 5.2.5. Factorial Schur functions. Pipedreams contain non-local information carried by the connectivity of the lines, just like the loop models of section 3. It is natural to wonder if there is a “vertex” representation similar to the six-vertex model for pipedreams. As this model is based on a Hecke algebra (which we have considered so far in the regular representation), there are in fact plenty 62 of vertex representations corresponding to various quotients of the Hecke algebra. The simplest one, analogous to the Temperley–Lieb quotient, is as follows. Suppose that σ is a Grassmannian permutation. By definition, this means that σ has at most one descent, i.e. there is a k such that i 6= k implies σ(i) < σ(i + 1). Then one can group lines into two subsets depending on whether they start on the vertical axis at position i ≤ k or i > k. If we use different colors for them, we get a picture such as 1 2 3 4 5 1 2 3 4 5 Several observations are in order. First, there is indeed no more non-local information in the sense that the connectivity of the endpoints can be entirely determined from the sequences of colors on the horizontal axis. Indeed, if the green endpoints are a1 < · · · < ak and the red endpoints are b1 < · · · < bn−k , there is only one Grassmannian permutation which is compatible with these colors, namely (a1 , . . . , ak , b1 , . . . , bn−k ). Secondly, there is never any crossing below row k, so Ξσ does not depend on xk+1 , . . . , xN . According to (5.5), we also know that it is a symmetric polynomial of x1 , . . . , xk . Thirdly, there are now five types of plaquettes: the four colorings of the non-crossing plaquette, but only one crossing plaquette (lines of the same color cannot cross each other!). In fact, by identifying red to occupied and green to empty, we recognize the six-vertex model configurations under the form of north-east going paths, in which the weight b2 = 0 (red paths cannot straight to the right). Taking into accounts the weights, this is exactly the free fermionic five-vertex model considered at the end of section 2.4.3 (Fig. 13). Furthermore, if we continue the red lines to the left so they all end on the same horizontal line, we find exactly the configurations that contribute to the Schur function sλ , where λ is encoded by the sequence of red and green endpoints at the top (which one can extend into an infinite sequence by filling with red dots at the left and green dots at the right). Finally, by comparing the weights, we find the equality: Ξσ (x1 , . . . , xn |0, . . . , 0) = sλ (x1 , . . . , xn ) In this case, the Ξσ (x1 , . . . , xn |y1 , . . . , yn ) are usually called factorial Schur functions. They are essentially the same as double Schur functions, see [81, 80] for a detailed discussion. 5.3. Orbital varieties. We shall move on to more sophisticated objects. In general, orbital varieties are the irreducible components of the intersection of a nilpotent orbit (by conjugation) with a Borel sub-algebra inside a Lie algebra. We cannot possibly reproduce the general theory of orbital varieties, and here, we shall restrict ourselves to a very special type of orbital varieties which corresponds to the Temperley–Lieb algebra (for more general orbital varieties from an “integrable” point of view, see [24, 25]). 5.3.1. Geometric description. We use the same notations as in section 5.2; but here all matrices are of even size N = 2n. Furthermore, call n+ = {strict upper triangular matrices} ⊂ b+ . Consider the affine scheme ON = {M ∈ n+ : M 2 = 0} 63 that is upper triangular matrices that square to zero. We have the following description of its irreducible components: they are indexed by link patterns π ∈ P2n . To each π we associate the upper triangular matrix π< with entries (π< )ij equal to 1 if i < j and (i, j) paired by π, 0 otherwise. For example, 0 0 0 0 0 0 0 1 .. 1 0 0 0 0 0 0 0 0 0 0 0 0 1 0 π= π< = 1 0 0 .0. 00 1 2 3 4 5 6 7 8 0 Then the corresponding irreducible component Oπ is given by the equations i i Oπ = {M ∈ g : M 2 = 0 and rank M (5.7) i j ≤ rank π< j , i, j = 1, . . . , N } where M j is the sub-matrix of M below and to the left of (i, j). Alternatively, Oπ can be defined as an orbit closure: Oπ = B+ · π< (5.8) B+ acts by conjugation Like all orbital varieties, ON is equidimensional, and the codimension of Oπ in n+ is n(n − 1). There is a torus T = (C× )N +1 acting on n+ , by conjugation by diagonal matrices and by scaling. The corresponding generators of its dual are denoted by x1 , . . . , xN and a. In fact, since conjugation by a scalar is trivial the torus acting is really of dimension N , and this amounts to saying that all multidegrees we shall consider only depend on differences xi − xj (and on a). Finally, the Oπ are T -invariant; we define Ωπ , the Joseph–Melnikov polynomial, to be Ωπ = mdegn+ Oπ These are extended Joseph polynomials, in the sense that the original Joseph polynomials correspond to a = 0 (no scaling action). Melnikov is the one that initiated the study of these particular orbital varieties [75]. The specialization a = 1, xi = 0, corresponding to considering the action by scaling only, gives the degree of Oπ (i.e. the number of intersections with a generic linear subspace of complementary dimension; in the case of a complete intersection, it is just the product of degrees of defining equations). Examples: • If π = 0n , the base link pattern, then O0n is a linear subspace where the upper-right n × n block is free while all other entries are zero. Thus, Y Y 0 ⋆ (5.9) O0n = Ω0n = (a + xi − xj ) (a + xi − xj ) 0 0 1≤i<j≤n We note the similarity with (4.10). • A random example in size N = 6: (5.10) π= Oπ = 1 2 3 4 5 6 n+1≤i<j≤2n ( M ∈ n+ : ) M12 = M34 = M35 = M45 = 0 (M 2 )16 = (M 2 )26 = 0 Again, there are simple rules to determine which equations are actually needed. Here we have six equations left which is the codimension, so that Ωπ = (a + x1 − x2 )(a + x3 − x4 )(a + x3 − x5 )(a + x4 − x5 )(2a + x1 − x6 )(2a + x2 − x6 ) Compare with the example at the end of section 4.2.1. 64 5.3.2. The Temperley–Lieb algebra revisited. In general, Joseph polynomials are known to be related to representations of the corresponding Weyl group, in our case the symmetric group. Note that the latter is nothing but the limit q → ±1 of the Hecke algebra. For the particular nilpotent orbit we have considered, we find unsurprisingly a representation of the Temperley–Lieb quotient of the symmetric group. However our extra scaling action makes the story much more interesting and produces a connection to the rational qKZ equation. Consider the Temperley–Lieb algebra with parameter τ = 2 (or q = −1) acting as in section 3.2.5 on the space of link patterns. The rational R-matrix is of the form Ři (u) = where we recall that graphically, ei = (a − u) + u ei a+u , and the identity is 1 = . It can be deduced from the trigonometric R-matrix (4.1) by sending q → −1 in the following way q = −e−~a/2 , z = e~u , ~→0 The rational qKZ system is (5.11) (5.12) Ři (xi − xi+1 )ΩN (x1 , . . . , xN ) = Ω(x1 , . . . , xi+1 , xi , . . . , xN ) ρ−1 ΩN (x1 , . . . , xN ) = (−1)n−1 ΩN (x2 , . . . , xN , x1 + 3a) It can be obtained from the (trigonometric) qKZ system (4.2,4.3) by once again sending q to −1: q = −e−~a/2 , zi = e−~xi , ~→0 The claim is that ΩN , the vector of Joseph–Melnikov polynomials defined in the previous section, solves (5.11,5.12), and in fact coincides with the ~ → 0 limit of (−1)n(n−1)/2 ~−n(n−1) ΨN , where ΨN is the solution of the trigonometric qKZ system that was discussed in section 4.2. We can proceed analogously to section 4.2.1 and rewrite (5.11) in components by separating it into two cases: Ωπ (i, i + 1) not paired in π : ∂i (5.13) =0 a + xi − xi+1 X (i, i + 1) paired in π : −(a + xi − xi+1 )∂i Ωπ = (5.14) Ωπ ′ π ′ 6=π,ei ·π ′ =π We shall now discuss the geometric meaning of (5.13,5.14). Since we know the base case (5.9), this will suffice to prove that ΩN , the vector of multidegrees, satisfies the whole qKZ system. Still, it would be satisfactory to have a geometric interpretation of (5.12) too. Unfortunately, it is currently unknown. Note that the effect of the r.h.s. of (5.12) on multidegrees can be quite drastic: for example, starting from the example (5.10), one goes back to the base case 03 , cf (5.9); but so doing, two quadratic equations have turned into linear equations! Remark: one can write a rational qKZ equation which is analogous to (4.5) but with additive spectral parameters. It should not be confused with the Knizhnik–Zamolodchikov (KZ) equation: the latter is recovered by the further limit a → 0, turning the difference equation into a differential equation. 65 5.3.3. The Hotta construction. The Hotta construction [37] is intended to explain the Joseph representation (Weyl group representation on Joseph polynomials) for an arbitrary orbital variety. When extended to our scaling action, it will produce the exchange relation (5.11). For more details see [25, 59]. The idea, as in the Schubert case, is to try to “sweep” with a subgroup Li , here acting by conjugation, and to apply the lemma of section 5.2.4. We need to be a little careful that our embedding space, n+ , is not Li -invariant, so we must translate multidegrees in n+ to multidegrees in g. It is easy to see that Y mdegg W = (a + xi − xj ) mdegn+ W W ⊂ n+ i≥j so that the effect of sweeping with Li amounts to a “gauged” divided difference operator ∂˜i : ∂˜i f = 1 a + xi+1 − xi ∂i ((a + xi+1 − xi )f ) = (a + xi − xi+1 )∂i f a + xi − xi+1 Now, letting Li and B+,i act by conjugation, start with an orbital variety Oπ which according to (5.8), is B+,i -invariant. If one tries to sweep it with Li , one can in general produce non-upper triangular matrices. In fact, a small calculation shows that this occurs exactly if Mi i+1 6= 0. So, we must distinguish two cases: • If (i, i + 1) are not paired in π, this means that the rank condition at (i, i + 1) in (5.7) says that Mi i+1 = 0; which in turn implies that a + xi − xi+1 | Ωπ . Furthermore, in this case Oπ is Li -invariant, so that ∂˜i Ωπ = 0. This is exactly (5.13). • If (i, i + 1) is a pair in π, then generically Mi i+1 6= 0 in Oπ . We then proceed in two steps. Cutting: first we intersect Oπ with the hyperplane Mi i+1 = 0. Since the intersection is transverse, the effect on the multidegree is to multiply by the weight of the hyperplane, which is a + xi − xi+1 . Sweeping: now we can sweep with Li and we stay inside ON . One can check that the map µ is generically one-to-one, and by dimension argument the image must be a union of orbital varieties (plus possibly some lower-dimensional pieces). The claim, which we shall not attempt to justify here, is that the orbital varieties thus obtained are exactly the proper preimages of π under ei . So we find −∂˜i ((a + xi − xi+1 )Ωπ ) = P π ′ 6=π,ei ·π ′ =π Ωπ ′ , which is equivalent to (5.14). 5.3.4. Recurrence relations and wheel condition. Several other constructions have simple geometric meaning. We mention in passing here the meaning of recurrence relations and of the wheel condition. Set xi+1 = xi + a. This gives the weight of 0 to Mi i+1 . Roughly, this corresponds to looking at what happens when Mi i+1 → ∞ (this is more or less clear from the integral formula of section 5.1.2; for a more precise statement, see [57]). To remain inside ON , writing the equations (M 2 )j i+1 = 0 and (M 2 )ij = 0, one concludes that one must have Mji = 0 and Mi+1 j = 0 for all j. At this stage one notes that the entries Mj i+1 and Mij are unconstrained by the equations. Removing the ith and (i + 1)st rows and columns reduces ON to ON −2 . What we have just shown is that when Mi i+1 → ∞, being in ON amounts to setting a certain number of entries to zeroes and then the result is some irrelevant linear space times ON −2 . One can be a bit more careful and do this reasoning at the level of each irreducible component: it is easy to see that only the components that have a pair (i, i + 1) survive (the others satisfy Mi i+1 = 0), 66 and that Oϕi π gets sent to Oπ . Finally, we get the recurrence relations: 0 π 6∈ Im ϕi 2n i−1 Y Y (2a + xi − xj ) (a + xj − xi ) (5.15) Ωπ (. . . , xi+1 = xi + a, . . .) = π = ϕi π ′ j=i+2 j=1 Ωπ′ (x1 , . . . , xi−1 , xi+2 , . . . , xN ) to be compared with (4.14). Similarly, if one sets xk = xj + 2a = xi + a, i < j < k, then the equation (M 2 )ik = 0 cannot possibly be satisfied since it contains the infinite term Mij Mjk , so all multidegrees must vanish, which is nothing but the wheel condition. 5.4. Brauer loop scheme. In this section we introduce a new affine scheme whose existence was suggested by the underlying integrable model. The subject has an interesting history, which we summarize now. In 2003, Knutson, in his study of the commuting variety, introduced the upper-upper scheme [55]; one of its components is closely related to the commuting variety, and in particular has the same degree. In an a priori unrelated development, de Gier and Nienhuis [18] studied the Brauer loop model, a model of crossing loops which is completely similar to the one that ) are allowed. They found that the we described in section 3.3.1, except crossing plaquettes ( entries of the ground state (or steady state of the corresponding Markov process) are again integers, and observed empirically that certain entries coincide with the degrees of the irreducible components of the upper-upper scheme. This mysterious connection required an explanation. A partial one was given in [26], where the corresponding inhomogeneous model was introduced and it was suggested that this generalization corresponds to going over from degrees to multidegrees. Also, an idea of the geometric action of the Brauer algebra was given. But some entries remained unidentified. In [58], the Brauer loop scheme was introduced, and it was proved that its top-dimensional components produce all the entries of the ground state. Finally, in the recent preprint [59], it is shown that more generally, an appropriate polynomial solution of the qKZ system associated to the Brauer algebra produces multidegrees of these components. 5.4.1. Geometric description. There are several equivalent descriptions of the Brauer loop scheme, see [58, 59]. Here we try to present it in a way which best respects the underlying symmetries. Let N be an integer. Consider complex upper triangular matrices that are infinite in both directions and that are periodic by shift by (N, N ): RZ mod N = {M = (Mij )i,j∈Z : Mij = Mi+N j+N ∀i, j ∈ Z} RZ mod N is an algebra. Let S = (δi,j−1 ) ∈ RZ mod N be the shift operator. Then we define the algebra b̂+ to be the quotient b̂+ = RZ mod N / S N b̂+ is finite-dimensional: dim b̂+ = N 2 . Inside b̂+ we have its radical n̂+ , which consists of (classes of) matrices with zero diagonal entries, and the group B̂+ of its invertible elements, i.e. with non-zero diagonal entries. The Brauer loop scheme is then defined as EN = {M ∈ n̂+ : M 2 = 0} For a different point of view on the origin of EN , see section 1.3 of [59]. 67 In all that follows, we take N to be even, N = 2n. We define a crossing link pattern to be an cr and is of cardinality (2n−1)!!. involution of Z/N Z without fixed points. Their set is denoted by P2n They have a similar graphical representation as link patterns, but in which crossings are allowed: 4 3 4 3 4 3 cr P4 = , , 1 2 1 2 1 2 The irreducible components of EN are known to be indexed by crossing link patterns [58, 94]. If cr , then π ∈ P2n Eπ = {M ∈ EN : (M 2 )i,i+N = (M 2 )j,j+N ⇔ i = j or i = π(j) mod N } Note that this definition (i) makes sense because (M 2 )i,i+N , despite being undefined for a generic element of b̂+ (it is killed by the quotient by S N ), is actually well-defined for elements of n̂+ and (ii) implies that the N numbers (M 2 )i,i+N always come in pairs for M ∈ EN , which is somewhat surprising (an elementary proof would be nice, as opposed to the proof of [58]). We can also describe Eπ as the closure of a union of orbits: (5.16) Eπ = B̂+ · tπ< B̂+ acts by conjugation, tπ< = {M ∈ b̂+ : Mij 6= 0 ⇒ i = π(j) mod N } Finally, there is a conjectural description in terms of equations: i (5.17) Eπ = {M ∈ b̂+ : M 2 = 0, (M 2 )i,i+N = (M 2 )π(i),π(i)+N , rank M i j ≤ rank π< j , j < i + N } where π< is any generic matrix in tπ< , for example the one made of zeroes and ones. Let us now turn to the torus action on b̂+ . First there is the usual scaling action, with generator a. But the remaining N -dimensional torus action is more difficult to explain. Let (xi )i∈Z be a set of formal variables satisfying the relations xi+N = xi + ǫ, where ǫ is another formal variable. Then the action of the full torus T = (C× )N +1 is defined by writing that wt(Mij ) = a + xi − xj i, j ∈ Z, i ≤ j It is slightly non-trivial but true that EN and the Eπ are T -invariant. We define the Brauer loop polynomials Υπ to be their multidegrees: Υπ = mdegn̂+ Eπ Υπ is a homogeneous polynomial in N + 1 variables, which we can choose to be a, xi − x1 (i = 2, . . . , N ) and ǫ. Its degree is the codimension of Eπ , which is 2n(n − 1). Examples: • The maximally crossed link pattern χn corresponds to χn (i) = i + n. In this case, Eχn is a linear subspace defined by the equations Mij = 0 if j ≤ i + n − 1. Thus, (5.18) n χ = Υ χn N i+n−1 Y Y (a + xi − xj ) = i=1 j=i+1 68 • At the opposite end, we find the non-crossing link patterns of P2n . Among them, we have 0n (or any of its rotations), which has all pairings parallel. In the Brauer loop scheme, they play a special role: Knutson proved, in the context of the upper-upper scheme [55], that there is a Gröbner degeneration of Cn × V to E0n , where V is some irrelevant vector space, and Cn is the commuting variety: Cn = {(X, Y ) ∈ gl(n, C) : XY = Y X} • Explicit examples can become quite complicated; even in size N = 4, we find, with the notation b = a − ǫ: 4 3 χ2 = Eχ2 = {M ∈ n̂+ : M12 = M23 = M34 = M45 = 0} 1 2 Υχ2 =(a + x1 − x2 )(a + x2 − x3 )(a + x3 − x4 )(b + x4 − x1 ) 4 3 02 = E02 1 2 M12 = M34 = 0 M23 M35 + M24 M45 = 0 = M ∈ n̂+ : M45 M57 + M46 M67 = 0 M13 M35 − M24 M46 = 0 Υ02 =(a + x1 − x2 )(a + x3 − x4 ) (a2 + ab + b2 − bx1 + ax2 + x1 x2 − ax3 − x2 x3 + bx4 − x1 x4 + x3 x4 ) 4 3 12 =ρ(02 ) = E12 1 2 Υ12 =(a + x2 − x3 )(b + x4 − x1 ) M23 = M45 = 0 M12 M24 + M13 M34 = 0 = M ∈ n̂+ : M34 M46 + M35 M56 = 0 M13 M35 − M24 M46 = 0 (a2 + 2ab + bx1 − ax2 − x1 x2 + ax3 + x2 x3 − bx4 + x1 x4 − x3 x4 ) The last two varieties are not complete intersections. 5.4.2. The Brauer algebra. The Brauer algebra is defined by generators fi , ei , i = 1, . . . , N − 1, and relations e2i = τ ei ei ei±1 ei = ei (5.19) fi2 = 1 (fi fi+1 )3 = 1 fi ei = ei fi = ei ei fi fi+1 = ei ei+1 = fi+1 fi ei+1 ei ej = ej ei fi fj = fj fi ei fj = fj ei ei+1 fi fi+1 = ei+1 ei = fi fi+1 ei |i − j| > 1 where indices take all values for which the identities make sense. Note that the fi are generators of the symmetric group SN , whereas the ei generate a Temperley–Lieb algebra. There is an associated solution of the Yang–Baxter equation, namely (5.20) Ři (u) = a(a − u) + a u ei + (1 − τ /2)u(a − u)fi (a + u)(a − (1 − τ /2)u) 69 The Brauer algebra acts naturally on crossing link patterns. The graphical rules are the same as in the previous sections, and we shall not illustrate them again. If π is viewed as an involution, then fi acts by conjugation by the transposition (i, i + 1), whereas ei creates new cycles (i, i + 1) and (π(i), π(i + 1)) – unless π(i) = i + 1, in which case it multiplies the state by τ . We now claim that the vector ΥN of multidegrees Υπ of the irreducible components of the Brauer loop scheme solves the qKZ system associated to the Brauer loop model [59], which we can write (5.21) Ři (xi − xi+1 )ΥN = si ΥN ∀i ∈ Z where si permutes xi+kN and xi+1+kN for all k, on condition that the following identification of the parameters is made: 2(a − ǫ) (5.22) τ= 2a − ǫ We could also add the cyclicity condition (which is obviously satisfied by ΥN by rotational invariance), but note that it could at most be marginally stronger than (5.21) (and in fact, it is not), because we have imposed (5.21) for all integer values of i. In other words, this is already an “affinized” version of the exchange relation because of our shifted periodicity property xi+N = xi +ǫ. For future use, let us write in components (5.21). We find the usual dichotomy: Υπ (5.23) = Υfi ·π π(i) 6= i + 1 : −(a + xi − xi+1 )((a + b)∂i + si ) a + xi − xi+1 X π(i) = i + 1 : −(a + xi − xi+1 )(a + b + xi+1 − xi )∂i Υπ = (a + b) (5.24) Υπ ′ π ′ 6=π,ei ·π ′ =π with the convenient notation b = a − ǫ. Remark: The Brauer algebra is the rational limit of the BWM algebra considered in [88]. 5.4.3. Geometric action of the Brauer algebra. This is the most technical part of [59], which we shall only sketch here. Once again, it follows the same general idea of trying to “sweep” with a subgroup L̂i analogous to the Li used so far. In our setting of infinite periodic matrices, this L̂i consists of invertible matrices which are the identity everywhere except in the entries Mab with a, b ∈ {i + kN, i + 1 + kN } for some k; and B̂+,i = L̂i ∩ B̂+ . The additional subtlety comes from the 2 fact that contrary N to the case of orbital varieties, the condition M = 0 is not stable by conjugation by L̂i (i.e. S is not stable by conjugation by L̂i , since the latter is outside b̂+ ). We shall have to reimpose one equation, namely (M 2 )i+1 i+N = 0, after sweeping. So, letting L̂i and B̂+,i act by conjugation, start with a component Eπ which according to (5.16), is B̂+,i -invariant. As usual we have to distinguish the two cases: • π(i) 6= i + 1: in this case we can directly sweep with L̂i , and then cut with (M 2 )i+1,i+N . As shown in [26, 58, 59] with varying levels of rigor and clarity, the result is precisely Eπ ∪Efi ·π . So we find, at the level of multidegrees, −(A + B + xi+1 − xi )∂˜i Υπ = Υπ + Υf ·π i which reduces after a few manipulations to (5.23). • π(i) = i + 1: this time we first cut with Mi i+1 = 0, sweep with L̂i (throwing away the L̂i -invariant piece which cannot contribute to the multidegree calculation), and then cut again with (M 2 )i+1 i+N . We lost one dimension in the process, S so the result cannot simply be a union of components. In fact, one can show that it is π′ 6=π,ei ·π′ =π Eπ ∩ {(M 2 )i i+N = (M 2 )i+1 i+N +1 }. This results directly in the multidegree identity (5.24). 70 Recurrence relations can also be written and interpreted geometrically as in section 5.3.4, but we shall omit them for brevity. Let us simply mention the wheel condition: the Υπ satisfy Υπ (xi = x, xj = x + a, xk = x + 2a) = 0 i < j < k < i+N (cf the appendix B of [88]). This stems geometrically from the equation (M 2 )ik = 0 (which is not killed by the quotient by S N since k < i + N ), which cannot be satisfied when Mij , Mjk → ∞. 5.4.4. The degenerate limit. There are two special values of the parameter τ . One, on which we shall not dwell, is τ = 1, which corresponds to ǫ = 0 in (5.22). This is the value for which ΥN becomes the ground state eigenvector of an integrable transfer matrix, that of [18], and is the subject of [26, 58]. Various interesting results can be obtained, including sum rules. In particular, in the homogeneous limit xi → 0 multidegrees become degrees and we recover the integer numbers observed in [18]. However, there is another special point: τ = 2. Indeed, we see that the coefficient of fi in the R-matrix (5.20) vanishes. In fact we recover this way the Temperley–Lieb solution of rational Yang–Baxter equation, which was used in connection with orbital varieties. What is the geometric meaning of this reduction? Plugging τ = 2 into (5.22), we note that there is no solution for ǫ, since ǫ → ∞ when τ → 2. Let us carefully take the limit ǫ → ∞ in the polynomials Υπ . We consider them as polynomials in z1 , . . . , zN , a and ǫ (this choice breaks the rotational invariance zi → zi+1 ) and keep the former fixed while sending the latter to infinity. Geometrically, the situation is as follows. Let us compute the weights of the entries Mij of a matrix M ∈ b̂+ in terms of z1 , . . . , zN , a and b = a − ǫ. Due to the periodicity, we can always assume i to be between 1 and N , and due to the quotient we should then consider i ≤ j < i + N . The result is that one finds two categories of entries: ( a + zi − zj 1≤i≤j≤N wt(Mij ) = b + zi − zj−N 1 ≤ j − N < i ≤ N This amounts to subdividing b̂+ (a vector space of dimension N 2 ) as U b̂+ = L where U is an upper triangular matrix, and L a strict lower triangular matrix. These two pieces have quite different destinies as τ → 2: the weights of U remain unchanged and are identical to those that we have used in section 5.3 for orbital varieties, whereas the weights of L diverge. It turns out that this limit corresponds to killing off this lower triangular part (as is clear from the definition of the multidegree as an integral, see section 5.1.2); and after taking the quotient, it is easy to see that we are left with b+ , the algebra of upper triangular matrices. Let us call p this projection (U, L) 7→ U from b̂+ to b+ . At the level of components, here is what happens: if π is a non-crossing link pattern, then p(Eπ ) = Oπ , the corresponding orbital variety. Translated into multidegrees, this means: B→∞ Υπ (z1 , . . . , zN , a, b) ∼ bn(n−1) Ωπ (z1 , . . . , zN , a) In general, crossing link patterns will project to lower dimensional B+ -orbit closures [59]. Interestingly, the matrix Schubert varieties and double Schubert polynomials of section 5.2 also appear as a special case in the τ = 2 limit. Indeed, consider the permutation sector, that is the 71 cr consisting of involutions such that π({1, . . . , n}) = {n + 1, . . . , 2n}. Such involutions subset of P2n are in bijection with permutations σ ∈ Sn , according to σ(i) = π(n + 1 − i) − n, i = 1, . . . , n: 6 1 2 3 7 7 5 σ= ←→ 1 2 3 6 4 8 π= 4 1 flip ←→ 8 5 1 4 π= 4 3 2 2 3 (the last flip is purely cosmetic and due to the fact that we have written labels of link patterns counterclockwise up to now). In the permutation sector, the structure of Eπ is as follows: 0 X ⋆ Eπ = 0 Y ⋆ If one discards the unconstrained entries (represented by ⋆), one recognizes the pairs of n × n matrices (X, Y ) in terms of which the upper-upper scheme of Knutson is defined [55]. In fact the union of Eπ where π runs over the permutation sector is exactly up to these irrelevant entries the upper-upper scheme. Now the projection of such components Eπ , that is here (X, Y ) 7→ X, is essentially the corresponding matrix Schubert variety Sσ . There are various ways to see that; one can prove it rigorously by using the description in terms of orbits; or one can compare the defining equations (5.1) of matrix Schubert varieties to the (conjectured) defining equations (5.17) of Eπ (which, in the case of the permutation sector, reduce to the first conjecture of section 3 of [55]). In any case, the result is more precisely that there is a vertical flip, so that p(Eπ ) ≃ Sσ σ0 . In terms of multidegrees, this means that Y Y B→∞ (a + xi − xj ) (a + xi − xj ) Υπ (x1 , . . . , x2n , a, b) ∼ bn(n−1)−|σ| 1≤i<j≤n n+1≤i<j≤2n Ξσ (a + xn , . . . , a + x1 |xn+1 , . . . , x2n ) (note that the prefactor is nothing but Ω0n ). So the exchange relation (5.3) satisfied by double Schubert polynomials (as well as the symmetric one (5.4)) should be somehow contained in the exchange relation (5.21) of Brauer loop polynomials. In order to see this, one must redefine the generator fi to ti = (1 − τ /2)fi when taking the limit τ → 2. The operators ti then satisfy the nil-Hecke algebra, and one can check that in the permutation sector, where the ei part of the Rmatrix never contributes when i 6= n, the exchange relation (5.21) reduces to (5.3) if i < n or to (5.4) if i > n. Acknowledgements I would like to thank A. Knutson, J.-M. Maillet, F. Smirnov who accepted to be members of my jury, and especially V. Pasquier, N. Reshetikhin and X. Viennot who agreed to referee this manuscript. I would also like to specially thank J.-B. Zuber, who besides collaborating with me on numerous occasions, has regularly advised in my work and provided guidance during the preparation of this habilitation. 72 I would like to thank my colleagues at my former laboratory (LPTMS Orsay) and new one (LPTHE Jussieu), especially their directors S. Ouvry and O. Babelon who welcomed me in their labs and were always open for discussions. My thanks to all my collaborators, here listed in the order of the number of co-publications (and alphabetically in case of ties): P. Di Francesco, with whom I have had the pleasure to have a most fruitful collaboration, which is the basis for a good part of the results in the manuscript; J. Jacobsen, J.-B. Zuber; V. Korepin, N. Andrei, J. de Gier, T. Fonseca, V. Kazakov, A. Knutson (the only person with whom I wrote a paper without ever meeting in person), P. Pyatov, A. Razumov, G. Schaeffer, Yu. Stroganov. My interaction with all of them has been influential in my scientific career. Finally, I would like to thank C. Le Vaou whose invaluable help with the administrative issues during all these years at LPTMS made it possible for me to concentrate on my research; and my family and friends for their support during the writing process. My work was supported by European Union Marie Curie Research Training Networks “ENRAGE” MRTN-CT-2004-005616, “ENIGMA” MRT-CT-2004-5652, European Science Foundation program “MISGAM” and Agence Nationale de la Recherche program “GIMP” ANR-05-BLAN0029-01. References [1] D. Allison and N. Reshetikhin, Numerical study of the 6-vertex model with domain wall boundary conditions, Ann. Inst. Fourier (Grenoble) 55 (2005), no. 6, 1847–1869, arXiv:cond-mat/0502314. MR MR2187938 [2] G. Andrews, Plane partitions. V. The TSSCPP conjecture, J. Combin. Theory Ser. A 66 (1994), no. 1, 28–39. MR MR1273289 [3] M. Batchelor, J. de Gier, and B. Nienhuis, The quantum symmetric XXZ chain at ∆ = −1/2, alternatingsign matrices and plane partitions, J. Phys. A 34 (2001), no. 19, L265–L270, arXiv:cond-mat/0101385. MR MR1836155 [4] R. Baxter, Exactly solved models in statistical mechanics, Academic Press, 1982. [5] A. Belavin, A. Polyakov, and A. Zamolodchikov, Infinite conformal symmetry in two-dimensional quantum field theory, Nucl. Phys. B 241 (1984), no. 2, 333, doi. [6] P. Bleher and V. Fokin, Exact solution of the six-vertex model with domain wall boundary conditions. Disordered phase, Comm. Math. Phys. 268 (2006), no. 1, 223–284, arXiv:math-ph/0510033. MR MR2249800 [7] P. Bleher and K. Liechty, Exact solution of the six-vertex model with domain wall boundary conditions. Critical line between ferroelectric and disordered phases, J. Stat. Phys. 134 (2009), no. 3, 463–485, arXiv:0802.0690. MR MR2485725 , Exact solution of the six-vertex model with domain wall boundary conditions. Ferroelectric phase, [8] Comm. Math. Phys. 286 (2009), no. 2, 777–801, arXiv:0712.4091. MR MR2472044 [9] D. Bressoud, Proofs and confirmations: The story of the alternating sign matrix conjecture, MAA Spectrum, Mathematical Association of America, Washington, DC, 1999. MR MR1718370 [10] Harish Chandra, Differential operators on a semi-simple Lie algebra, Amer. J. Math. 79 (1957), 87–120. [11] I. Cherednik, Quantum Knizhnik–Zamolodchikov equations and affine root systems, Comm. Math. Phys. 150 (1992), no. 1, 109–136, http://projecteuclid.org/getRecord?id=euclid.cmp/1104251785. MR MR1188499 [12] , Double affine Hecke algebras, London Mathematical Society Lecture Note Series, vol. 319, Cambridge University Press, Cambridge, 2005. MR MR2133033 [13] H. Cohn, M. Larsen, and J. Propp, The shape of a typical boxed plane partition, New York J. Math. 4 (1998), 137–165, arXiv:math/9801059. MR MR1641839 [14] F. Colomo and A. Pronko, The arctic circle revisited, 2007, arXiv:0704.0362. , Emptiness formation probability in the domain-wall six-vertex model, Nuclear Phys. B 798 (2008), [15] no. 3, 340–362, arXiv:0712.1524. MR MR2411855 [16] , The limit shape of large alternating sign matrices, 2008, arXiv:0803.2697. 73 [17] J. de Gier, Loops, matchings and alternating-sign matrices, Discrete Math. 298 (2005), no. 1-3, 365–388, arXiv:math/0211285. MR MR2163456 [18] J. de Gier and B. Nienhuis, Brauer loops and the commuting variety, J. Stat. Mech. Theory Exp. (2005), no. 1, 006, 10 pp, arXiv:math.AG/0410392. MR MR2114232 [19] J. de Gier and P. Pyatov, Factorised solutions of Temperley–Lieb qKZ equations on a segment, arXiv:0710.5362. [20] J. de Gier, P. Pyatov, and P. Zinn-Justin, Punctured plane partitions and the q-deformed Knizhnik– Zamolodchikov and Hirota equations, J. Combin. Theory Ser. A 116 (2009), 772–794, arXiv:0712.3584, doi. [21] D. Dei Cont and B. Nienhuis, The packing of two species of polygons on the square lattice, J. Phys. A 37 (2004), 3085–3100, arXiv:cond-mat/0311244. [22] P. Di Francesco, Totally symmetric self-complementary plane partitions and the quantum Knizhnik– Zamolodchikov equation: a conjecture, J. Stat. Mech. Theory Exp. (2006), no. 9, P09008, 14 pp, arXiv:cond-mat/0607499. MR MR2278472 [23] P. Di Francesco and P. Zinn-Justin, Around the Razumov–Stroganov conjecture: proof of a multi-parameter sum rule, Electron. J. Combin. 12 (2005), Research Paper 6, 27 pp, arXiv:math-ph/0410061. MR MR2134169 [24] , Quantum Knizhnik–Zamolodchikov equation, generalized Razumov–Stroganov sum rules and extended Joseph polynomials, J. Phys. A 38 (2005), no. 48, L815–L822, arXiv:math-ph/0508059, doi. MR MR2185933 [25] , From orbital varieties to alternating sign matrices, 2006, extended abstract for FPSAC’06, arXiv:math-ph/0512047. , Inhomogeneous model of crossing loops and multidegrees of some algebraic varieties, Comm. Math. [26] Phys. 262 (2006), no. 2, 459–487, arXiv:math-ph/0412031. MR MR2200268 [27] , Quantum Knizhnik–Zamolodchikov equation: reflecting boundary conditions and combinatorics, J. Stat. Mech. Theory Exp. (2007), no. 12, P12009, 30 pp, arXiv:0709.3410, doi. MR MR2367185 , Quantum Knizhnik–Zamolodchikov equation, totally symmetric self-complementary plane partitions and [28] alternating sign matrices, Theor. Math. Phys. 154 (2008), no. 3, 331–348, arXiv:math-ph/0703015, doi. [29] P. Di Francesco, P. Zinn-Justin, and J.-B. Zuber, A bijection between classes of fully packed loops and plane partitions, Electron. J. Combin. 11 (2004), no. 1, Research Paper 64, 11 pp, arXiv:math/0311220. MR MR2097330 [30] , Determinant formulae for some tiling problems and application to fully packed loops, Ann. Inst. Fourier (Grenoble) 55 (2005), no. 6, 2025–2050, arXiv:math-ph/0410002. MR MR2187944 [31] L. D. Faddeev, How Algebraic Bethe Ansatz works for integrable model, 1996, arXiv:hep-th/9605187. [32] S. Fomin and A. Kirillov, The Yang–Baxter equation, symmetric functions, and Schubert polynomials, Proceedings of the 5th Conference on Formal Power Series and Algebraic Combinatorics (Florence, 1993), vol. 153, 1996, pp. 123–143. MR MR1394950 [33] S. Fomin and R. Stanley, Schubert polynomials and the nil-Coxeter algebra, Adv. Math. 103 (1994), no. 2, 196–207. MR MR1265793 [34] T. Fonseca and P. Zinn-Justin, On the doubly refined enumeration of alternating sign matrices and totally symmetric self-complementary plane partitions, Electron. J. Combin. 15 (2008), Research Paper 81, 35 pp, arXiv:0803.1595. MR MR2411458 [35] I. Frenkel and N. Reshetikhin, Quantum affine algebras and holonomic difference equations, Commun. Math. Phys. 146 (1992), 1–60, http://projecteuclid.org/euclid.cmp/1104249974. [36] I. Gessel and G. Viennot, Binomial determinants, paths, and hook length formulae, Adv. in Math. 58 (1985), no. 3, 300–321. MR MR815360 [37] R. Hotta, On Joseph’s construction of Weyl group representations, Tohoku Math. J. (2) 36 (1984), no. 1, 49–74. MR MR733619 [38] C. Itzykson and J.-B. Zuber, The planar approximation. II, J. Math. Phys. 21 (1980), no. 3, 411–421. MR MR562985 [39] A. Izergin, D. Coker, and V. Korepin, Determinant formula for the six-vertex model, J. Phys. A 25 (1992), no. 16, 4315–4334. MR MR1181591 [40] A. G. Izergin, Partition function of a six-vertex model in a finite volume, Dokl. Akad. Nauk SSSR 297 (1987), no. 2, 331–333, http://adsabs.harvard.edu/abs/1987SPhD...32..878I. MR MR919260 [41] J. Jacobsen and J. Kondev, Field theory of compact polymers on the square lattice, Nucl. Phys. B 532 (1998), 635–688, arXiv:cond-mat/9804048. [42] , Conformal field theory of the Flory model of polymer melting, Phys. Rev. E 69 (2004), no. 6, 066108, arXiv:cond-mat/0209247, doi. [43] J. Jacobsen and P. Zinn-Justin, Algebraic Bethe Ansatz for the FPL2 model, J. Phys. A 37 (2004), no. 29, 7213–7225, arXiv:math-ph/0402008, doi. MR MR2078954 74 [44] M. Jimbo and T. Miwa, Solitons and infinite-dimensional Lie algebras, Publ. Res. Inst. Math. Sci. 19 (1983), no. 3, 943–1001, http://projecteuclid.org/euclid.prims/1195182017. MR MR723457 [45] , Algebraic analysis of solvable lattice models, CBMS Regional Conference Series in Mathematics, vol. 85, Published for the Conference Board of the Mathematical Sciences, Washington, DC, 1995. MR MR1308712 [46] W. Jockush, J. Propp, and P. Shor, Random domino tilings and the arctic circle theorem, arXiv:math.CO/9801068. [47] K. Johansson, The arctic circle boundary and the Airy process, Ann. Probab. 33 (2005), no. 1, 1–30, arXiv:math/0306216. MR MR2118857 [48] A. Joseph, Orbital varieties, Goldie rank polynomials and unitary highest weight modules, Algebraic and analytic methods in representation theory (Sønderborg, 1994), Perspect. Math., vol. 17, Academic Press, San Diego, CA, 1997, pp. 53–98. MR MR1415842 [49] M. Kasatani, Subrepresentations in the polynomial representation of the double affine Hecke algebra of type GLn at tk+1 q r−1 = 1, Int. Math. Res. Not. (2005), no. 28, 1717–1742, arXiv:math/0501272. MR MR2172339 [50] M. Kasatani and Y. Takeyama, The quantum Knizhnik–Zamolodchikov equation and non-symmetric Macdonald polynomials, Funkcial. Ekvac. 50 (2007), no. 3, 491–509, arXiv:math/0608773. MR MR2381328 [51] P. Kasteleyn, Graph theory and crystal physics, Graph Theory and Theoretical Physics, Academic Press, London, 1967, pp. 43–110. MR MR0253689 [52] D. Kazhdan and G. Lusztig, Representations of Coxeter groups and Hecke algebras, Invent. Math. 53 (1979), no. 2, 165–184, doi. MR MR560412 [53] R. Kenyon, Lectures on dimers, http://www.math.brown.edu/~ rkenyon/papers/dimerlecturenotes.pdf. [54] A. Kirillov, Jr. and A. Lascoux, Factorization of Kazhdan–Lusztig elements for Grassmannians, Combinatorial methods in representation theory (Kyoto, 1998), Adv. Stud. Pure Math., vol. 28, Kinokuniya, Tokyo, 2000, pp. 143–154, arXiv:math.CO/9902072. MR MR1864480 [55] A. Knutson, Some schemes related to the commuting variety, J. Algebraic Geom. 14 (2005), no. 2, 283–294, arXiv:math.AG/0306275. MR MR2123231 [56] A. Knutson and E. Miller, Gröbner geometry of Schubert polynomials, Ann. of Math. (2) 161 (2005), no. 3, 1245–1318. MR MR2180402 [57] A. Knutson, E. Miller, and A. Yong, Gröbner geometry of vertex decompositions and of flagged tableaux, Journal für die reine und angewandte Mathematik (2007), arXiv:math.CO/0502144. [58] A. Knutson and P. Zinn-Justin, A scheme related to the Brauer loop model, Adv. Math. 214 (2007), no. 1, 40–77, arXiv:math.AG/0503224. MR MR2348022 [59] , The Brauer loop scheme and orbital varieties, 2009, preprint. [60] V. Korepin, Calculation of norms of Bethe wave functions, Comm. Math. Phys. 86 (1982), no. 3, 391–418. MR MR677006 [61] V. Korepin and P. Zinn-Justin, Thermodynamic limit of the six-vertex model with domain wall boundary conditions, J. Phys. A 33 (2000), no. 40, 7053–7066, arXiv:cond-mat/0004250, doi. MR MR1792450 [62] G. Kuperberg, Another proof of the alternating-sign matrix conjecture, Internat. Math. Res. Notices (1996), no. 3, 139–150, arXiv:math/9712207. MR MR1383754 [63] A. Lascoux, Symmetric functions and combinatorial operators on polynomials, CBMS Regional Conference Series in Mathematics, vol. 99, Published for the Conference Board of the Mathematical Sciences, Washington, DC, 2003. MR MR2017492 [64] A. Lascoux and M.-P. Schützenberger, Polynômes de Kazhdan & Lusztig pour les grassmanniennes, Young tableaux and Schur functors in algebra and geometry (Toruń, 1980), Astérisque, vol. 87, Soc. Math. France, Paris, 1981, pp. 249–266. MR MR646823 [65] G. Lawler, O. Schramm, and W. Werner, Values of Brownian intersection exponents. I. Half-plane exponents, Acta Math. 187 (2001), no. 2, 237–273. MR MR1879850 [66] , Values of Brownian intersection exponents. II. Plane exponents, Acta Math. 187 (2001), no. 2, 275–308. MR MR1879851 , Values of Brownian intersection exponents. III. Two-sided exponents, Ann. Inst. H. Poincaré Probab. [67] Statist. 38 (2002), no. 1, 109–123. MR MR1899232 [68] E. Lieb, Exact solution of the F model of an antiferroelectric, Phys. Rev. Lett. 18 (1967), no. 24, 1046–1048, doi. [69] , Exact solution of the problem of the entropy of two-dimensional ice, Phys. Rev. Lett. 18 (1967), no. 17, 692–694, doi. [70] , Exact solution of the two-dimensional Slater KDP model of a ferroelectric, Phys. Rev. Lett. 19 (1967), no. 3, 108–110, doi. , Residual entropy of square ice, Phys. Rev. 162 (1967), no. 1, 162–172, doi. [71] 75 [72] B. Lindström, On the vector representations of induced matroids, Bull. London Math. Soc. 5 (1973), 85–90. MR MR0335313 [73] I. Macdonald, Affine Hecke algebras and orthogonal polynomials, Astérisque (1996), no. 237, Exp. No. 797, 4, 189–207, Séminaire Bourbaki, Vol. 1994/95. MR MR1423624 [74] P. MacMahon, Combinatory analysis, Cambridge University Press, 1915. [75] A. Melnikov, B-orbits in solutions to the equation X 2 = 0 in triangular matrices, J. Algebra 223 (2000), no. 1, 101–108. MR MR1738254 [76] E. Miller and B. Sturmfels, Combinatorial commutative algebra, Graduate Texts in Mathematics, vol. 227, Springer-Verlag, New York, 2005. MR MR2110098 [77] W. Mills, D. Robbins, and H. Rumsey, Jr., Alternating sign matrices and descending plane partitions, J. Combin. Theory Ser. A 34 (1983), no. 3, 340–359. MR MR700040 [78] W. Mills, D. Robbins, and H. Rumsey, Jr, Self-complementary totally symmetric plane partitions, J. Combin. Theory Ser. A 42 (1986), no. 2, 277–292. MR MR847558 [79] T. Miwa, On Hirota’s difference equations, Proc. Japan Acad. Ser. A Math. Sci. 58 (1982), no. 1, 9–12, http://projecteuclid.org/euclid.pja/1195516178. MR MR649054 [80] A. Molev, Comultiplication rules for the double Schur functions and Cauchy identities, 2008, arXiv:0807.2127. [81] A. Molev and B. Sagan, A Littlewood–Richardson rule for factorial Schur functions, Transactions of the American Mathematical Society 351 (1999), no. 11, 4429–4443, arXiv:q-alg/9707028. [82] B. Nienhuis, Phase transitions and critical phenomena, vol. 11, Academic Press, 1987, Eds. C. Domb and J. Lebowitz. [83] J.D. Noh and D. Kim, Interacting domain walls and the five-vertex model, arXiv:cond-mat/9312001. [84] I. Nolden, The asymmetric six-vertex model, J. Stat. Phys. 67 (1992), 155–201, doi. [85] S. Okada, Enumeration of symmetry classes of alternating sign matrices and characters of classical groups, J. Algebraic Combin. 23 (2006), no. 1, 43–69, arXiv:math/0408234, doi. MR MR2218849 [86] A. Okounkov and N. Reshetikhin, Correlation function of Schur process with application to local geometry of a random 3-dimensional Young diagram, J. Amer. Math. Soc. 16 (2003), no. 3, 581–603, arXiv:math.CO/0107056. MR MR1969205 [87] V. Pasquier, Scattering matrices and affine Hecke algebras, Low-dimensional models in statistical physics and quantum field theory (Schladming, 1995), Lecture Notes in Phys., vol. 469, Springer, Berlin, 1996, pp. 145–163, arXiv:q-alg/9508002. MR MR1477945 [88] , Incompressible representations of the Birman–Wenzl–Murakami algebra, Ann. Henri Poincaré 7 (2006), no. 4, 603–619, arXiv:math/0507364. MR MR2232366 [89] , Quantum incompressibility and Razumov Stroganov type conjectures, Ann. Henri Poincaré 7 (2006), no. 3, 397–421, arXiv:cond-mat/0506075. MR MR2226742 [90] A. Razumov and Yu. Stroganov, Combinatorial nature of the ground-state vector of the O(1) loop model, Teoret. Mat. Fiz. 138 (2004), no. 3, 395–400, arXiv:math/0104216, doi. MR MR2077318 [91] , Bethe roots and refined enumeration of alternating-sign matrices, J. Stat. Mech. Theory Exp. (2006), no. 7, P07004, 12 pp, arXiv:math-ph/0605004. MR MR2244324 [92] A. Razumov, Yu. Stroganov, and P. Zinn-Justin, Polynomial solutions of qKZ equation and ground state of XXZ spin chain at ∆ = −1/2, J. Phys. A 40 (2007), no. 39, 11827–11847, arXiv:0704.3542, doi. MR MR2374053 [93] N. Reshetikhin, Lectures on the six-vertex model, Les Houches lecture notes. [94] B. Rothbach, Equidimensionality of the Brauer loop scheme, 2009, preprint. [95] J. Shore and D. J. Bukman, Coexistence point in the six-vertex model and the crystal shape of fcc materials, Phys. Rev. Lett. 72 (1994), no. 5, 604–607, doi. [96] F. Smirnov, A general formula for soliton form factors in the quantum sine–Gordon model, J. Phys. A 19 (1986), L575–L578. [97] R. Stanley, A baker’s dozen of conjectures concerning plane partitions, Combinatoire énumérative (Montreal, Que., 1985/Quebec, Que., 1985), Lecture Notes in Math., vol. 1234, Springer, Berlin, 1986, pp. 285–293. MR MR927770 [98] Yu. Stroganov, Izergin–Korepin determinant at a third root of unity, Teoret. Mat. Fiz. 146 (2006), no. 1, 65–76, arXiv:math-ph/0204042. MR MR2243403 [99] B. Sutherland, Exact solution of a two-dimensional model for hydrogen-bonded crystals, Phys. Rev. Lett. 19 (1967), no. 3, 103–104, doi. [100] B. Wieland, A large dihedral symmetry of the set of alternating sign matrices, Electron. J. Combin. 7 (2000), Research Paper 37, 13 pp, arXiv:math/0006234. MR MR1773294 [101] F. Wu, Remarks on the modified potassium dihydrogen phosphate model of a ferroelectric, Phys. Rev. 168 (1968), no. 2, 539–543, doi. 76 [102] D. Zeilberger, Proof of the alternating sign matrix conjecture, Electron. J. Combin. 3 (1996), no. 2, Research Paper 13, 84 pp, The Foata Festschrift, arXiv:math/9407211. MR MR1392498 [103] , Proof of a conjecture of Philippe Di Francesco and Paul Zinn-Justin related to the qKZ equations and to Dave Robbins’ two favorite combinatorial objects, 2007, http://www.math.rutgers.edu/~ zeilberg/mamarim/mamarimhtml/diFrancesco.html. [104] P. Zinn-Justin, The influence of boundary conditions in the six-vertex model, arXiv:cond-mat/0205192. [105] , Six-vertex model with domain wall boundary conditions and one-matrix model, Phys. Rev. E 62 (2000), no. 3, part A, 3411–3418, arXiv:math-ph/0005008. MR MR1788950 , HCIZ integral and 2D Toda lattice hierarchy, Nuclear Phys. B 634 (2002), no. 3, 417–432, [106] arXiv:math-ph/0202045. MR MR1912027 [107] , Proof of the Razumov–Stroganov conjecture for some infinite families of link patterns, Electron. J. Combin. 13 (2006), no. 1, Research Paper 110, 15 pp, arXiv:math/0607183. MR MR2274325 , Combinatorial point for fused loop models, Comm. Math. Phys. 272 (2007), no. 3, 661–682, [108] arXiv:math-ph/0603018, doi. MR MR2304471 [109] , Littlewood–Richardson coefficients and integrable tilings, Electron. J. Combin. 16 (2009), Research Paper 12, arXiv:0809.2392. Paul Zinn-Justin, LPTMS (CNRS, UMR 8626), Univ Paris-Sud, 91405 Orsay Cedex, France; and LPTHE (CNRS, UMR 7589), Univ Pierre et Marie Curie-Paris6, 75252 Paris Cedex, France. E-mail address: pzinn @ lpthe.jussieu.fr 77
© Copyright 2025 Paperzz